1

A Review of Millimeter Wave Device-based

Localization and Device-free Sensing
Technologies and Applications
Anish Shastri, Student Member, IEEE, Neharika Valecha, Student Member, IEEE,
Enver Bashirov, Student Member, IEEE, Harsh Tataria, Member, IEEE, Michael Lentmaier, Senior Member, IEEE,
Fredrik Tufvesson, Fellow, IEEE, Michele Rossi, Senior Member, IEEE, and Paolo Casari, Senior Member, IEEE
arXiv:2112.05593v3 [cs.NI] 25 May 2022

Abstract—The commercial availability of low-cost millimeter- Device-free (i.e., radar-based) sensing systems still have to be
wave (mmWave) communication and radar devices is starting to improved in terms of: improved accuracy in the detection of vital
improve the adoption of such technologies in consumer markets, signs (respiration and heart rate) and enhanced robustness/gen-
paving the way for large-scale and dense deployments in fifth- eralization capabilities across different environments; moreover,
generation (5G)-and-beyond as well as 6G networks. At the same improved support is needed for the tracking of multiple users,
time, pervasive mmWave access will enable device localization and for the automatic creation of radar networks to enable large-
and device-free sensing with unprecedented accuracy, especially scale sensing applications. Finally, integrated systems performing
with respect to sub-6 GHz commercial-grade devices. joint communications and sensing are still in their infancy:
This paper surveys the state of the art in device-based local- theoretical and practical advancements are required to add sens-
ization and device-free sensing using mmWave communication ing functionalities to mmWave-based channel access protocols
and radar devices, with a focus on indoor deployments. We based on orthogonal frequency-division multiplexing (OFDM)
overview key concepts about mmWave signal propagation and and multi-antenna technologies.
system design, detailing approaches, algorithms and applications Index Terms—Millimeter waves; propagation characteristics;
for mmWave localization and sensing. Several dimensions are channel models; communications; localization; sensing; radar;
considered, including the main objectives, techniques, and per- practical constraints;
formance of each work, whether they reached an implementation
stage, and which hardware platforms or software tools were used. I. I NTRODUCTION
We analyze theoretical (including signal processing and ma-
chine learning), technological, and implementation (hardware Millimeter-wave (mmWave) communications in the 28–
and prototyping) aspects, exposing under-performing or missing 300 GHz band are looked at with great interest, as they
techniques and items towards enabling a highly effective sensing may be able to quench –at least temporarily– the ever-
of human parameters, such as position, movement, activity increasing bandwidth requirements of such applications as
and vital signs. Among many interesting findings, we observe
that device-based localization systems would greatly benefit from
massive Internet of things (IoT), virtual/augmented reality,
commercial-grade hardware that exposes channel state informa- mobile cloud services and ubiquitous ultra-high definition mul-
tion, as well as from a better integration between standard- timedia streaming [1]–[3]. This would cover the shortcomings
compliant mmWave initial access and localization algorithms, of sub-6 GHz technologies such as WiFi and fourth-generation
especially with multiple access points (APs). Moreover, more (4G) cellular networks, which currently cannot support the
advanced algorithms requiring zero-initial knowledge of the envi-
ronment would greatly help improve the adoption of mmWave si-
massive bandwidth and number of users the above applications
multaneous localization and mapping (SLAM). Machine learning imply.
(ML)-based algorithms are gaining momentum, but still require The potential of mmWave technology, however, is not
the collection of extensive training datasets, and do not yet limited to higher-rate communications: rather, mmWave de-
generalize to any indoor environment, limiting their applicability. vices can become a proxy for high-resolution device-based
localization as well as device-free sensing. These capabilities
Manuscript received xxxx xx, xxxx . . . follow from the physics of mmWave propagation. First, the
This work received support from the European Commission’s Horizon 2020 shorter wavelength of mmWaves (compared to sub-6 GHz
Framework Programme under the Marie Skłodowska-Curie Action MINTS signals) enables accurate location estimates and lower location
(GA no. 861222), and from Italian Ministry for Education, University and
Research (MIUR) under the “Departments of Excellence” initiative (Law error bounds [4], [5]. Second, mmWaves have well-known
232/2016). and peculiar propagation characteristics [6], [7] which yield
A. Shastri (email: anish.shastri@unitn.it) and P. Casari (email: higher spatial scanning resolution. For example, mmWaves
paolo.casari@unitn.it) are with the Department of Information Engineering
and Computer Science, University of Trento, 38123 Povo (TN), Italy. propagate quasi-optically, meaning that a line-of-sight (LoS)
N. Valecha (email: neharika.valecha@eit.lth.se), M. Lentmaier multipath component (MPC) is predominant over non-line-of-
(email: michael.lentmaier@eit.lth.se) and F. Tufvesson (email: sight (NLoS) contributions to the received signal [8]. Scatter-
fredrik.tufvesson@eit.lth.se) are with the Department of Electrical and
Information Technology, Lund University, 22100 Lund, Sweden. ing also has a limited impact off typical non-rough reflecting
E. Bashirov (email: enver.bashirov@dei.unipd.it), and M. Rossi (email: surfaces such as walls, furniture, metal plates as well as glass
rossi@dei.unipd.it) are with the Department of Information Engineering, layers [9], [10].
University of Padova, 35131 Padova, Italy.
H. Tataria (email: harsh.tataria@ericsson.com) was with Ericcson AB, Another consequence of mmWave propagation is that
22363 Lund, Sweden, when working on this paper. mmWave signals undergo much higher path loss with respect
 2

to microwaves. To compensate for this attenuation, and still on mmWave signal structure and propagation characteristics
enable long-reach wireless links, mmWave devices resort to that make this domain unique with respect to other radio
large or massive antenna arrays. Via beamforming, they can communication and sensing technologies. We consider practi-
focus their transmitted energy towards a confined portion of cal constraints that define the applicability of algorithms and
the 3D space, and thus achieve greater directionality. While processing schemes to mmWave devices operating indoors.
this requires specific protocols for initial access [11]–[13] We then delve into a detailed description of device-based
and beam training such as the IEEE 802.11ad [14], [15] and indoor localization algorithms, explaining the main localiza-
802.11ay [16], [17] standards, it also means that a reduced tion techniques employed in the literature, and how they are
amount of power is typically directed towards secondary practically implemented in real mmWave hardware whenever
multipath components. In addition with the quasi-optical available. For device-free sensing, we list a number of relevant
propagation patterns discussed above, the main consequence applications and technologies that leverage mmWave hardware
is that the received angular spectrum of a mmWave signal and signals to detect, localize and track targets indoors, as well
is sparse: in typical conditions, one can identify one LoS as to specifically identify features related to sub-sections of a
MPC along with a number of NLoS MPCs corresponding target (e.g., a part of the human body). Because these device-
to signal reflections off the surrounding environment. The free approaches are mainly based on mmWave radar devices,
above features of mmWave communications have significant we will briefly discuss how mmWave radar bands are being
implications for localization and sensing [18]. For example, standardized for different applications.
being able to separate MPCs in the angular domain enables
angle-based localization schemes that are not normally used
A. Differences with respect to previous surveys
in sub-6 GHz systems due to limited angular resolution when
using small antenna arrays. Fingerprinting-based algorithms Localization and sensing are topics of great interest for both
can also be enhanced by incorporating angle-based features current and future-generation wireless communication system
to improve location discrimination. From the point of view engineering. The research on these topics has proceeded at
of device-free sensing, mmWave propagation also implies a steady pace, considering aspects as diverse as localization
typically clearer reflections off sensed targets and parts thereof. techniques, heterogeneous technologies, different scenarios,
For example, a) quasi-optical mmWave propagation along with and different kinds interactions between the device to be lo-
b) the large mmWave bandwidth available at typical mmWave calized and the location server, among others. Several surveys
radar frequencies respectively imply that reflections off targets cover these aspects, typically for sub-6 GHz technologies. For
are usually separate in the a) angle and b) time domains. example, Zafari et al. [23] and Geok et al. [24] focus on
This makes it possible to measure features that point to each localization techniques for wireless systems in general, and
reflection’s movement velocity (e.g., the Doppler shift) and use cover heterogeneous technologies. These works only tangen-
this data to precisely localize and identify different targets. tially consider mmWaves, and instead survey geometric and
In this paper, we focus on indoor mmWave device-based signal processing-based localization methods for sub-6 GHz
localization and device-free sensing, and provide a com- systems. Ngamakeur et al. [25] delve into device-free sensing
prehensive review of approaches, technologies, schemes and of different human signatures using sub-6 GHz technologies
algorithms to estimate a device or object’s location in an indoors. Here, the focus is on the localization, tracking and
indoor environment. The objective of our survey is to shed identification of multiple subjects using Wi-Fi and other kinds
light on indoor applications of localization and sensing using of wireless sensors.
mmWave signals. Location information can be extremely By leveraging similar technologies, Singh et al. [26] con-
useful in different indoor setups [19], [20]. For example, in sider techniques and algorithms to localize IoT devices in-
factories and industrial environments, location information can doors. In this case, the focus of the survey is on a specific
be exploited to enhance ultra-reliable low-latency communi- source of location information (received WiFi signal strength
cations (URLLC) for industrial IoT and smart manufactur- fingerprints) and on how machine learning works when applied
ing [21], [22]. Accurate localization and sensing can benefit to such datasets. By expanding into the concept of smart world,
healthcare scenarios for patient tracking and lifesign/behavior the work in [27] also surveys how sub-6 GHz technologies
monitoring, help people navigate in indoor areas, provide can help improve a variety of services via data collection
trajectory suggestions through relevant waypoints in museums, and system automation using active and passive sensing tech-
malls, and company headquarters, as well as support mission- niques. Finally, the work in [28] touches on aspects related
critical applications such as disaster relief and indoor security. to the modeling and estimation of wireless channels in fifth-
Location systems are also crucial for network performance generation (5G) cellular systems. While the work touches on
optimization. Accurate location information can support the localization, the covered techniques apply to outdoor cellular
fast alignment of transmit and receive antenna arrays, optimize systems, and can thus leverage the density and much higher
the association between clients and access points (APs), and computational power of their hardware.
prevent blockage of high-power LoS paths via predictive Unlike our survey, none of the above works targets mil-
handovers to provide seamless coverage. This can result in limeter wave device-based and device-free indoor localization.
low-latency communications as needed for augmented reality, This area is characterized by several interesting research works
virtual reality, and tactile Internet applications. to date, but remains a very hot topic due to the inception of
In the following, we start with an overview that touches mmWave coverage for future 5G-and-beyond networks as well
 3

as wireless (indoor) local-area networks. The objective of our specular reflections depend on the electrical thickness of the
survey is to cover the most significant work in this area, while wall, which in turn is also a function of frequency. Interest-
giving a comprehensive view of unsolved challenges and open ingly, we have no evidence that the reflection coefficient varies
research avenues. with frequency, although the transmission power decreases
Note that, in our survey, we are not seeking an analysis uniformly with increasing frequency due to the skin effect in
of the limits of mmWave localization and sensing technology lossy media [36].
based on purely theoretical arguments, or an operational Two effects that have gained spotlight with the increased
description of well-known geometric localization algorithms, interest in the mmWave band are diffraction and diffuse
or even a coverage of the integration between mmWave scattering. The former reduces noticeably at high frequencies,
communications and 5G, beyond-5G, and future 6G networks. and larger objects lead to “sharp” shadows. The latter effect is
These are related yet tangential topics for which we rather refer more significant as the surface roughness becomes comparable
the interested reader to one of the several excellent surveys that to mmWave wavelengths. As the surface roughness increases,
touch on these aspects, e.g., [18], [19], [22], [28]–[34]. the objects behave like a Lambertian radiator, which scatters
the radiation. Foliage has a similar effect as scattering; with
B. Outline and organization of the manuscript the decreasing wavelength relative to the size of the leaves, we
observe more diffused scattering and less penetration. Another
The remainder of this paper expresses three purposes: to
factor is atmospheric attenuation due to fog or rain [37]
cover the characteristics of mmWave propagation and commu-
and may affect the mmWave frequencies in case of extreme
nication/sensing hardware that impacts localization and sens-
weather.
ing performance, including standardization efforts (Sections II
through IV); to detail the state of the art in device-based Channel models used for localization need to account for
mmWave localization (Section V) and in device-free mmWave the above mentioned phenomena, and are often based on
sensing (Section VI); and finally to discuss our findings, ray tracing or cluster-based modeling with some geometry-
discuss promising research avenues, and draw concluding based stochastic channel model (GSCM) [38]–[40]. Moreover,
remarks (Sections VII and VIII). for ray tracing approaches, high-resolution environment in-
In particular, Sections V and VI constitute the core of formation is needed to account for such surface roughness,
our technological survey. Section V discusses device-based as different materials have different properties (e.g., glass
localization algorithms for indoor environments, whereas Sec- windows vs. concrete walls). These effects also depend on the
tion VI presents several approaches for radar-based device- environment: the high concrete walls and glass surfaces of the
free localization. Each section is organized to first present urban areas lead to different propagation conditions, compared
the section topic, and then to add progressively more details to the greener suburban areas with, e.g., stucco exteriors and
related to the typical techniques appropriate for each section, shorter walls.
the hardware typically used in testbeds, and the description
of each surveyed approach. We also include summary tables B. Measurement techniques and results
to help the reader navigate the contents and extract key
To model the properties of a channel, we need to perform the
information. Both Sections V and VI end with a summary
measurements for different propagation scenarios. A channel
of the most relevant aspects and findings.
sounder, that helps to measure these properties is not only
Fig. 1 represents the organization of the survey as a mind
an expensive piece of equipment but as we move towards
map, starting from Section II (top right), proceeding clockwise,
higher frequencies, the susceptibility to phase noise as well
and concluding with Section VII.
as antenna spacing errors start to increase. Similarly, the cost
and energy consumption of up/down-conversion chains, in
II. I NFLUENCE OF MM WAVE CHANNELS
particular of the front-end mixed signal circuitry in analog-to-
A. Impact of mmWave frequencies on propagation conditions digital and digital-to-analog converters (ADCs/DACs) as well
The propagation of a wave through any medium depends on as power amplifiers (PAs) becomes of paramount importance.
its frequency: this basic property helps us predict the behavior For up-to-Gbit/s sampling rates (as often required by best-in-
of the channel for diffeangularrent carrier frequencies. When class channel sounding), 12-15 bit resolution is required. To
it comes to mmWaves, considering the Friis equation under penetrate larger distances (and thus to maximize the forward
the assumption that the antenna gain G at both link ends is link gain), PAs typically need to operate with 6-10 dB backoff
frequency-independent (by reducing the antenna aperture), the power efficiency and need to be continuously driven close to
free space path loss increases with the square of the carrier their 1 dB compression point limits.
frequency f . On the contrary, assuming a constant physical Consequently, the channel sounders used often for measure-
area A at both the transmitter (TX) and the receiver (RX), ments at high frequencies use omnidirectional antennas [41]
the antenna gains G = A(4π/λ)2 increase on both sides, or if directional [42], then the angular resolution is not
and thus the overall path loss decreases quadratically with in- taken into account. Directionality is achieved by mechanically
creasing frequency f [35]. Specular reflections for dielelectric rotating horn antennas in most cases and the angular resolution
halfspaces (e.g., ground reflections) depend on frequency as corresponds to the beamwidth, e.g., [43]–[45]. For indoor
long as the dielectric constant is itself a function of frequency. measurement scenarios, the directional information though can
For reflections at a dielectric layer (e.g., building walls) the be enhanced by using switched antenna arrays along with
 4

Measurement
techniques
§II-B

Open research Impact of

propagation
Selected directions §II-A
applications §VII
Channel
§VI-F
models
§II-C

Main learning mmWave

techniques channel §II
§VI-E

Radar-enabled
localization and
sensing §VI Summary §II-D

Key processing
techniques
mmWave Indoor
§VI-D Localization Analog
and Sensing §III-A

Beamforming
architectures
§III

Hybrid
approaches Indoor Performance
Hybrid
§V-G localization vs Complexity
§III-B
§III-D
algorithms §V-A Measurements
§V-B

Digital
Standardization §III-C
§IV
RSSI and ToF
Algorithms
§V-F
Evaluation
tools
§V-C

CSI Angle-based
§V-E §V-D

Fig. 1. Mind map showing the organization of this survey.

super-resolution algorithms like space-alternating generalized On the other hand, the variance of the path loss around
expectation maximization (SAGE) [46] and RIMAX [47]. It the distance-dependent mean is higher at mmWave frequen-
is possible to use electronically-switched horn arrays [48] as cies, which in turn increases the probability of outage [51].
well, which additionally lets us evaluate the MPC and intra- The standard deviation as well is strongly dependent on the
cluster information. distance and its values increases from 5-10 dB to more than
1) Key outdoor results: When it comes to outdoor measure- 20 dB as the distance increases from 30 m to 200 m [50]. This
ments, path loss is a key parameter. For channel modelling, is due to the variation in power levels caused by location and
we need to measure the pathloss coefficient, its mean and its orientation of a street in an urban macro cell [52] and not due
variance. The pathloss coefficient for mmWave frequencies to shadowing as one may expect.
is close to that of microwaves, i.e., often there is no strong Another parameter important for channel modelling is the
frequency dependence beyond the f 2 dependence of free-space root mean square (RMS) delay spread. But it changes with
path loss [49]. In LoS scenarios, the path loss coefficient lies frequency and thus it may not be the best parameter to model
between 1.6-2.1 (2 for pure free-space propagation) and in the delay dispersion. Instead, delay windows may be a better
NLoS scenarios the value increases to 2.5 and 5 (e.g., [43], alternative as they define the time interval containing part of
[44], [50]). the energy of power-delay profile (PDP). Delay spreads in an
 5

TABLE I
S UMMARY OF CHANNEL MODELS AND THEIR SPATIAL PARAMETER VALUES

mmWave Channel Models

Parameter 3GPP [59] / COST METIS QuaDRiGa NYUSIM
ITU-R [60] IRACON [61] [53] [62] [63]
f (GHz) 6 2.6 0.45-63 5.4 28
Type 2D GSCM GSCM 3D Map-based & GSCM 3D GSCM TCSL
K- factor µK 7 N/A 7.9 -1.6 N/A
σK 4 N/A 6 2.7 N/A
Delay Spread µDS -7.7 1.07 -7.42 -7.22 2.7
σDS 0.18 0.93 0.32 0.08 1.4
µASA 1.62 3.94 1.65 1.67 19.3
AOA Spread
σASA 0.22 3.91 0.47 0.15 14.5
µASD 1.60 0.71 1.64 1.54 23.5
AOD Spread
σASD 0.18 0.59 0.43 0.1 16.0
µZSA 1.22 3.73 1.28 1.61 7.4
ZOA Spread
σZSA 0.297 2.11 0.26 0.07 3.8
µZSD N/A 1.95 1.31 1.17 -7.3
ZOD Spread
σZSD N/A 1.80 0.31 0.07 3.8
µXP R 11 15.59 29 13 N/A
XPR (dB)
σXP R 4 10.39 6.5 1.6 N/A
µP L 47.9 N/A N/A 36.1 N/A
Shadow fading
σP L 3 N/A 3 1.6 N/A

outdoor environment are measured or simulated by ray tracing than 5 ns in LoS conditions, and 10–20 ns in NLoS condi-
[43]–[45], [53]. Beamforming can help with minimizing the tions [65], [70]–[73]. Though these measurements were lim-
delay spread [54]. The type of beamforming to be used ited to under 100 GHz, recently [74] performed measurements
depends on the angular dispersion properties. Angular spreads at 142 GHz and observed delay spread values of 3 ns in LoS
measured at the base station (BS) are more accurate than those and 9 ns for NLoS. Further, the observed channels are much
measured at the user equipment (UE) as the ray tracers used sparser at frequencies over 100 GHz and we notice higher
often do not include scattering objects such as street signs, partition loss compared to 28 GHz. It is worth noting for
parked cars, etc. in their geographic database [50], [55]. As indoor measurements, the number of MPCs is higher with
observed in [44], [56], the RMS angular spread at the BS is more clusters than measured for outdoor with rotating horn
of the order of 10◦ with one cluster only while at the UE, the antennas [75]. Here the angular spreads are often measured
angular spreads are in the range 30-70◦ [43], [44], [56], [57]. for clusters, with the intra-cluster azimuth and elevation angles
More information related to fixed wireless scenarios can be are described as having a Laplacian distribution with a spread
found in [58]. of 5◦ [76].
2) Key indoor results: Measurements for indoor environ-
ments have picked up in recent years as we look at localization
applications for 5G. The results are often from office and C. Models for mmWave channels
industrial environments, where different material densities can Because mmWave propagation channels differ from mi-
be studied. The path loss coefficient in this case ranges crowave channels, we need to redefine or rather add certain
from 1.2-2 in LoS to 2-3 in NLoS scenarios [64], [65]. The parameters for mmWave channel modeling. As mentioned
frequency dependence of the path loss is more significant in [58], mmWave channels require 3D modeling of azimuth
for indoor than outdoors, f k with k ≈ 2.5 was observed as well as elevation spreads, inclusion of temporal/spatial/fre-
in [66]. Overall though, the values are closer to those at sub- quency consistency and multipath cluster based modeling.
6 GHz, with an increased probability of outage. Path loss These have further impact when we consider positioning and
in some cases is shown to follow a dual-slope model and localization. Prevalent models for mmWave are GSCMs that
is the same for both mmWave and sub-6 GHz. The floating imitate the propagation environment with stochastic processes,
intercept model is another alternative used in Third-generation and create a 3D map. To correctly reproduce the wireless
partnership project (3GPP) standards for indoor modelling environment, parameter values need to be extracted from the
at high frequencies. Human blockage can cause upto 10- channel impulse response of real time measurements done
20dB attenuation regardless of one or two people [67] and using a channel sounder. An extensive review of propagation
similar values in case of trucks in outdoor scenarios [68]. chacteristics at mmWave frequencies is available in [77],
In [69], fast Fourier transform (FFT) based beamforming is which also provides a summary of channel sounder mea-
used in conjunction with a very large virtual array (25×25×25 surements and relevant channel models. The 3GPP defined
elements). It highlights the scattering caused by small objects different environments for mmWave channel modeling, these
specifically in NLoS case and the importance of small scale include Urban Macro, Urban Micro, Indoor Office and Rural
characterization. Further, it is shown that the indoor environ- Macro. Several outdoor and indoor measurements are avail-
ment leads to enhanced diffused MPC energy. able, but for this paper we compare large-scale parameter
Delay spread measured in office scenarios is usually less values for an indoor office scenario listed in Table I.
 6

Fig. 2. Cross-polarized antenna array panel [59].

Prominent channel models have been developed for the

Fig. 3. BS antenna array pattern as a function of azimuth and elevation scan
above mentioned scenarios based on measurements done in angles [81].
each of them. Some key results have already been discussed,
but we also observe that cluster-based multipath channel com-
ponents have been modelled, in order to specifically account communications, it is paramount that the channel evolves
for an indoor office environment. Also, as can be seen from the smoothly without discontinuities during mobility [81].
table, the angular spread is no longer limited to the azimuth 4) Polarization: The radiation pattern of each antenna
plane. element of an array extends over both the azimuthal plane
1) Static vs. dynamic modeling: Due to the high frequency and the elevation plane, and should be separately modelled
and thus higher path loss, there is significant deterioration for directional performance gains. Moreover, as we consider
when the UE is stationary and more so when the UE is indoor scenarios with higher number of reflections, the polar-
moving or is in a high movement zone and transitions from ization properties of the multipath components also come into
a LoS to NLoS scenario. This requires the dynamic modeling play.
of the communication channel, as the moving objects in 5) Large bandwidth and large antenna arrays: Antenna
the vicinity also act as random blocking obstacles. The BS arrays that are larger in size and also massive in the number
needs to transmit training beams more frequently so as to of antenna elements are needed at mmWave, thus high reso-
update the angle of departure (AoD)/angle of arrival (AoA) lution channel modeling includes propagation patterns both
estimates, since the location of UE changes over time, and in the angular domain and in the delay domain. Massive
slight errors in the orientation of the beams can lead to MIMO channel models [82] have previously not considered
significant performance loss [78]. So far, we have considered these exceptions but at mmWave, accurately modeling of the
a fixed BS and slow moving UE, but with 5G and vehicle- higher number of multipath components and their AoA/AoD is
to-everything (V2X) communications we expect high mobility paramount. Antenna elements in azimuth and elevation plane
scenarios [79]. Most mmWave channel models are still defined both need to be evaluated to consider all possible array struc-
only for a fixed BS, but have added support for dynamic tures (planar array, rectangular array, cylindrical array). Fig. 2
modeling scenarios for V2X. depicts an antenna array panel used for 3GPP/International
2) Blockage: mmWaves cannot penetrate obstacles such as telecommunication union – radiocommunication Sector (ITU-
human bodies, walls, foliage, etc. Thus, these blockage sources R) antenna modeling [59], [60]. Figs. 3 and 4 show the BS
need to be modelled in the link budget itself. One such charac- and UE array radiation pattern based on parameters as defined
terization study is found in [80], which measured power loss in [59, Table 7.3-1, page 22].
(in dB) when 70-GHz mmWave signals propagate through a
brick wall, a PC monitor, and book shelves. Blockage does not D. Summary
affect just the total received power but also the angle or power The mmWave channel when considered for indoor appli-
of multipath signal components, due to varying sizes, positions cations differs from the microwave channel in key aspects
and directions of the blocking object/human. Localizing the such as free space path loss, diffraction, and penetration loss
position of the UE with respect to these blockage sources with respect to different surfaces. This required the need
becomes onerous, especially in a dynamic setting. for different measurements to be done for channel charac-
3) Spatial consistency and clusters: A new, previously terization. Some key results are presented in Section II-B.
unexplored requirement was added to 3GPP Release 14 [59]. Path loss equations and penetration loss for indoor scenarios
When mmWave communications take place through narrow can be found in [59, Tables 7.4.1-1 and 7.4.3-1]. Various
antenna radiation beams, the channel characteristics become channel models have been developed, these include those by
highly correlated, especially when two UEs are close and 3GPP [59], ITU-R [60], METIS [53], MiWEBA [45], Fraun-
see the same BS. Also, for applications related to V2X hofer HHI’s QuaDRiGa [62], COST2100 [83], NYUSIM [84]
 7

and cost- efficient. Each phase shifter multiplies its input

2πk √
by ej 2N , where j = −1, N is the number of bits, and
k = 0, . . . , 2N − 1 is used to control the phase shifters.
Most commonly, codebook-based schemes are used to steer
the beams in the direction of the UE/receiver. At the re-
ceiver, the received signal strength indicator (RSSI) is the
most commonly used parameter to estimate the direction of
arrival and delay, and thus localize the device. However, phase
shifters have a constant amplitude constraint and limited phase
resolution. It is also worth noting that analog beamforming
converges to a single beam for multiple data transmissions,
and in multi-user case the inter-user interference is very high.
This is a drawback for localization applications, as the phase
resolution for analog beamforming is low.
The popularity of analog beamforming systems comes from
the availibility of commercial off-the-shelf (COTS) devices,
Fig. 4. UE antenna array pattern as a function of azimuth and elevation scan that are being used for research on mmWave positioning. The
angles [81]. devices come with a pre-programmed codebook to generate
beam patterns and with support for retrieving the RSSI and
channel state information (CSI) which can be used to isolate
which still has ongoing measurements for indoor scenarios. the position of the UE. One such hardware front-end is avail-
The channel models are all GSCM-based with added cluster able from TMYTEK, an analog correlator with beamformer
based modeling. Small-scale parameter values are further chips and smart-antenna arrays [89]. Another company that
available when considering indoor scenarios found in the provides beamformer integrated circuits and scalable anten-
documentations mentioned for corresponding models. nas for mmWave is Anokiwave [90]. Siver Semiconductors
Several measurements have been done in the mmWave band provides transceiver modules for mmWave frequencies, i.e.,
for outdoor (urban macro and urban micro) scenarios but the 28 GHz and 60 GHz [91]. National Instruments (NI) also has
indoor measurements are limited to the sub-6 Ghz band for the PXIe-5831, a mmWave vector signal transceiver that has
the channel models developed with the exception of [85], beamforming capabilities and phased antenna arrays [92]. It
where the authors propose an extension for an indoor channel has been used for channel measurements as mentioned above
model based on extensive measurements carried out at 28 as well [93]. We discuss the hardware devices used in more
and 140 GHz. We observe that indoor channel models are detail in Section V-C1.
an extension of outdoor ones, and can be adapted easily based
on the delay and angular spreads of any environment, as well
as by adapting path loss modeling. B. Hybrid beamforming
Hybrid beamforming is by far the most researched form
III. I MPLICATIONS OF BEAMFORMING ARCHITECTURES of beamforming, as it provides a middle ground between
FOR MM WAVE LOCALIZATION complexity and cost. Here, the analog beamformer is used
It is a common misconception that for higher frequencies in the RF domain, along with a digital precoder at baseband.
the free space propagation loss is higher. As explained in [86], This can be either a fully connected structure or a partially
[87], for given aperture area of the antennas used, shorter connected one. Hybrid analog/digital beamforming structures
wavelengths propagate farther due to the narrow directive provide balance between the beam resolution and cost and
beams. This is further verified in [88] with a patch antenna power consumption. By using multiple RF chains concurrently,
operated at 3 GHz and an antenna array operated at 30 GHz beam sweeping can be done in a short time leading to shorter
of the same physical size. We observe equal amounts of beam training time which leads to higher effective data rate. At
propagation loss irrespective of the operating frequency. Thus, mmWave frequencies the sparse channel behaviour is useful
mmWave frequencies enable the use of antenna arrays that for beam training and higher array gains. Multiple hybrid
produce highly directional beams which lead to large array beamforming techniques for mmWave have been proposed in
gains. This can be observed from Fig. 5, which shows not only the last ten years which broadly fall under codebook depen-
the increase in array size with respect to the beam penetration dent, spatially sparse precoding, antenna selection and beam
distance, but also how the larger array size increases the selection [94]. [95] first gave the idea of what we call hybrid
coverage area [35]. beamforming today. It was a combination of a digital baseband
precoder and an RF precoder which falls under spatially sparse
precoding. The work in [96] first proposed the idea of base-
A. Analog beamforming band beamforming, or “hybrid beamforming” as the authors
Analog beamforming, sometimes also referred to as beam named it, that chooses the best RF beam based on a capacity
steering, is done by connecting a single radio frequency (RF) maximization criterion, and then derives a zero-forcing (ZF)-
chain to a string of phase shifters that are both energy- based weighing matrix for digital precoding. Also, both [97]
 8

30 GHz Array Sizes

Arrival distance of 1 e

0m 20 cm
0m

Beam Arrival Distances

85 m
300 m 300 m 20 x 20
40 cm
600 m 4x4 600 m 40 x 40

900 m 900 m 80 x 80
8x8

1200 m 1200 m 80 cm
16 x 16
Array Sizes

Fig. 5. Effect of beamwidth relative to operating frequency and array sizes [35].

and [98] suggest codebook-based precoding solutions. Recent frameworks to do digital beamforming for a mmWave setup
works have proposed compressive sensing, least squares- and using linearization to help with power amplifier loss and
discrete Fourier transform (DFT)-based solutions for hybrid improved quantization.
beamforming with use cases in car-to-car scenarios and high
speed trains. In most cases, hybrid beamforming is seen to D. Performance vs. complexity overview
perform as well as fully-digital beamforming, and as being
both cost-effective and spectrally efficient. In localization applications, the requirement for mmWave
indoor systems is to isolate the position of the receiver
inside a room, while taking into account blockage caused
C. Digital beamforming by humans and objects alike, with LoS being the dominant
Digital beamforming adjusts the amplitude and phase of component. The presence of pillars, metal and glass surfaces
the transmitted signals using precoding. Linear precoding affects the channel impulse response and thus make it difficult
algorithms such as matched filter (MF), ZF, and regularized to extract position information. Presence of antenna arrays
zero-forcing (RZF) methods were classically used for single- greatly enhances the accuracy of the position coordinates.
antenna user systems. For multiple-antenna users, block diag- Whereas digital systems have cleaner isolated beams and can
onalization is a feasible approach. Digital beamforming can be potentially yield centimeter-level pointing accuracy, analog
considered as the best option for mmWave positioning. With setups have a limit to the number of beam patterns they
the possibility of huge antenna arrays (256 × 128 upwards) a can generate: when trying to increase the resolution, these
beam resolution of the order of centimeters can be achieved. beam patterns eventually start to overlap. As stated above, the
The calibration accuracy of digital systems allows us to number of beams is proportional to the number of available RF
use high-resolution parameter estimation algorithms that can chains, thus increasing the complexity hundred-fold for digital
estimate not only the time of arrival (ToA) and AoA but also systems. Calibration issues also prevent analog systems from
the Doppler frequency offset in case of mobility, making it performing high-resolution parameter estimation which could
possible to update the position of a UE in real-time. The issue improve the localization accuracy. Hybrid beamforming seems
here arises from the use of a RF chain per antenna, which leads a promising tradeoff as of now, due to the easier availability
to a complex, non-cost-effective hardware system for massive of COTS devices, and to a performance almost as good as that
multiple-input multiple-output (MIMO) structures. of fully digital systems.
As digital beamforming offers higher beam resolution, it
is a viable candidate where multi user mmWave or rather IV. P ROGRESS IN STANDARDIZATION OF CELLULAR
mmWave massive MIMO systems are considered. However, MM WAVE SYSTEMS
commercial hardware for a fully digital system is still in its The frequency bands used for 5G systems were proposed
infancy, and only laboratory results exist. Several authors have at the 2015 World Radio Conference (WRC) by ITU-R and
proposed alternative techniques for the realization of a digital approved during WRC 2019. The frequency bands standard-
system that is power efficient. For instance [99] gives an option ized by 3GPP in Release 15-17 [104]–[106] for 5G systems
for digital beamforming that employs switches to bypass the are classified as FR-I region (below 7.125 GHz) and FR-
hardware constraint of using multiple RF chains. In [100]– II region (between 7.25 GHz and 71 GHz). The approved
[102], the authors propose different ways to form an antenna FR-II bands are (in GHz): 24.25–27.5; 31.8–43.5; 45.5–
array using waveguides and printed circuit boards that support 50.2; 50.4–52.6; 66–71. FR-I bands act as the key bands for
digital beamforming. Alternatively, [99], [103] propose novel cellular communications, while the FR-II are more suited to
 9

(SRSs) with Release 16 extensions added uplink-based local-

ization or BS/AP centric localization as shown in Fig. 7b.
In this case, the UE sends the reference signal. Based on
the received SRSs, the BSs/APs can measure and report (to
the location server) the arrival time, the received power and
the AoAs from which the position of the UE is estimated.
The time difference between downlink reception and uplink
transmission can also be reported, and used in round-trip time
(RTT)-based positioning schemes, where the distance between
a BS/AP and a UE can be determined based on the estimated
RTT. By combining several such RTT measurements, involv-
ing different BSs/AP anchors, it becomes possible to estimate
the location of the UE.
We note that these methods do not utilize the full-
Fig. 6. 3GPP Release 16 radio access type-dependent architecture standard- dimensional nature of the propagation channel (azimuth and
ized for UE localization in URLLC scenarios. All BSs/APs are interfaced elevation domains), and do not fully take into account the
with a centralized unit enroute to a URLLC core network.
phase information needed to estimate the underlying MPCs
with high resolution. While this is an ongoing topic for
short-range communications. The FR-II bands also provide research in many study items of 3GPP Releases 17 and 18,
increased bandwidths compared to FR-I, and are managed we refer the reader to [106], [107] for further details. Along
via licensed access mechanisms such as enhanced UTRA- this same line, a steady stream of work is also conducted in
dual connectivity (EN-DC). As some bands overlap with other academia, see e.g., [108].
services, coexistence management is needed for terrestrial
access in overlapping satellite communication channels and V. D EVICE - BASED MM WAVE LOCALIZATION ALGORITHMS
for fronthaul and backhaul in fixed wireless systems. 5G com- FOR INDOOR COMMUNICATION SYSTEMS
mercial deployments have already been taking place since the
A. Introduction
end of last year, and some spectrum congestion was observed
initially amongst multiple operators. Since then, some novel In this section, we introduce algorithms and methods that
forms of spectrum access/coordination mechanisms have been leverage lab-grade and commercial-grade mmWave hardware
implemented.1 When it comes to localization of UEs, it was to localize devices indoors. We start with a brief recap on
the focus of 3GPP Release 16 [105] especially for the use case classical methods for indoor radio localization. The standard
of URLLC. In the past, Global Navigation Satellite Systems techniques designed for localization involve exploiting the pa-
assisted by cellular networks have been mostly used for UE rameters of radio signals from existing wireless infrastructure.
positioning, but their accuracy is high only in outdoor environ- These have been well explored and surveyed in, e.g., [29],
ments, as they rely on satellites to localize UEs. As we move [109]–[114]. With reference to Fig. 8, localization algorithms
towards higher frequencies, we require localization indoors typically make use of signal parameters related to received
as well, and we can accomplish it in 5G networks using the signal power (RSSI and signal-to-noise ratio, SNR), time-
location server, as it was for long-term evolution-advanced information such as time of flight (ToF) and time difference-
(LTE-A) systems. The location server collects and provides of-arrival (TDoA), and angle information (AoA and AoD) in
position estimates and assistance data and measurements to order to obtain distance and direction estimates, which enable
the other devices. Various localization methods are used, based a device or group of devices to estimate either their own
on downlink or uplink communications, either separately or in location, or the location of another device in their proximity, or
combination, to meet the accuracy requirements for different both. Fig. 9 offers a general view of this process considering
scenarios. The overall architecture is as depicted in Fig. 6. the papers on mmWave localization surveyed in the literature.
As shown in Fig. 7a, downlink-based localization is per- After a device has extracted location-dependent features from
formed when each of multiple BSs/APs send a different a received signal, such features are either used directly for
reference signal, known as the positioning reference signal localization, or further processed to extract additional infor-
(PRS). The UE receives the different PRSs and reports the ToA mation, or joined into a global map of the environment along
difference for PRSs received from multiple distinct BSs/APs with other measurements. The device then applies geometric
to the location server. The location server can use the reports or machine learning (ML)/deep learning (DL) algorithms to
to determine the position of the UE. Compared to LTE- derive location information.
Advanced, the PRS has a more regular structure and a much The most typical localization techniques rely on geometric
larger bandwidth, which enables a more precise correlation algorithms. For example, trilateration and triangulation utilize
and ToA estimation. distance and angle measurements from fixed reference points
The canonical 3GPP Release 15 sounding reference signals to compute an intersection, which yields the estimate of a
device’s location [110]. The reference points are usually the
1 We note that these are operator- and vendor-specific, since frequency band location of the access points, and the localized device is
combinations vary depending on the specific country. typically a client. The distances between the APs and the client
 10

are measured by exploiting either the ToF of the signal or we remark that wireless devices typically collect location-
by mapping the RSSI information to absolute distance using dependent signal features through the interaction between a
path-loss models. Fig. 10a shows an illustration of trilateration client and one or more APs. Such interactions naturally take
using ToF to estimate distances. place in mmWave networks, e.g., during standard-compliant
AoA (the angle at which the received signal strikes the link establishment and beam refinement procedures (see also
receiver antenna or antenna array) and angle difference-of- Section V-C). Therefore, in principle the measurement of
arrival (ADoA) (the difference between two AoAs), are es- signal features does not require the devices to implement
timated by applying signal parameter estimation algorithms localization-specific message exchange protocols. This makes
(like multiple signal classification (MUSIC) [115] and estima- localization an almost-inherent feature of mmWave communi-
tion of signal parameters via rotational invariance techniques cation systems [30], [118].
(ESPRIT) [116]) on the received signal. The AoAs from As remarked in Section III, however, mmWave devices
different APs are then triangulated to localize the client device. have peculiar characteristics that differentiate them from com-
Fig. 10b illustrates the triangulation-based technique, whereas monplace WiFi equipment. Specifically, mmWave arrays can
Fig. 10c depicts ADoA-based localization. incorporate a large number of antennas. The presence of large
Wireless channel characteristics, e.g., in the form of the arrays enable mmWave devices to output low-level physical
channel impulse response (CIR) between a transmitter and layer measurements from each antenna separately. Once the
a receiver, also provide valuable information for localization device has locked onto a signal, each antenna receives the
purposes, including the ToF of the received signal. The CSI same signal with a different phase, corresponding to the
can also be extracted from the receiver antennas to obtain delay incurred by the signal due to its spatial position in the
rich information about multipath signal components [117]. array. These measurements can be made available as CSI and
As a result, one can separate the LoS propagation path from localization algorithms can exploit them to localize a device,
NLoS paths, or detect that only NLoS components reached the either by converting them into AoA estimates (e.g., [119],
receiver, thus improving the accuracy of the signal parameter [120]) or by directly inferring the location of a device by
measurements. exploiting the CSI as a location-dependent feature.
The advent of bandwidth-hungry applications such as aug- Whenever CSI measurements are not available, a device can
mented reality, virtual reality, etc., and the ever-increasing still retrieve angle information by post-processing the output of
demand for high data rates, has made mmWave communi- standard-compliant beam training procedures. Typically, each
cation technology a popular potential replacement for existing mmWave has a number of pre-programmed beam patterns
WLAN systems. This is mainly due to the availability of large that provide it with the necessary flexibility to focus energy
bandwidth in the frequency range of 30-300 GHz, resulting in towards different directions. Each beam pattern ideally covers
multi-Gbit/s data rates. mmWaves propagate quasi-optically, a well-defined portion of the 3D space, so that observing each
thus reflecting crisply off indoor surfaces and obstacles with beam pattern separately makes it possible to implement a scan
limited scattering just like light rays [6]. This makes finer of all azimuthal and elevation angles that the mmWave array
measurements of signal parameters such as RSSI, AoA, can cover. Therefore, measuring the power received through
SNR, and ToF, more feasible and more accurate. Moreover, each beam pattern configuration would implement a sweep

BS2 ref

BS1 ref BS2 UE ref

UE ref
BS1

Location
Server
Location
Server UE
UE ref
UE UE ref

BS3 ref

BS4 ref

BS3 BS4

(a) Downlink based, UE-centric (b) Uplink based, BS/AP-centric

Fig. 7. Architecture of BS/AP-centric vs UE-centric localization.
 11

propagation path at each antenna. Post-processing CSI yields

Surface RSSI / SNR different signal parameters, including path attenuation and
angle information. If CSI values are sufficiently precise
Client device
(e.g., no coarse quantization affects the amplitude or phase),
Time
collecting receiver-side CSI from multiple antennas also
enables the estimation of AoAs. We cover CSI-based
AoA CIR / CSI
approaches in Section V-E.
Pros: Rich information that can be readily used for ranging
or as an input to learning-based approaches.
Time Cons: Typically not available straightforwardly on all devices.
ToF
Different devices may provide different types of CSI.
AoD

AP RSSI [132]–[139] — RSSI is one of the simplest proxies for

the range of a device in an environment. It is measured at a
Fig. 8. Illustration of the signal measurements obtained from mmWave receiving device as the power or amplitude of the received RF
propagation. The color gradient of the beam represents the decreasing signal
strength due to path loss.
signal. mmWave received signal strength (RSS) measurements
can be extracted from the physical or medium access control
(MAC) layer of a device and used to measure the distance of
of lookout angles. By identifying the beam pattern that leads a client from the AP, based on the knowledge of a path loss
to the largest received power, a mmWave device could easily model. The client is believed to lie on the circumference of
estimate angles of arrival. We now proceed to discuss each the circle centered on the AP and having the estimated range
type of location-dependent feature separately in the context of as the radius. Such estimates from more than two APs can
mmWave communications, highlighting the pros and cons of be trilaterated to approximate the location of the client. We
each feature. cover RSSI-based approaches in Section V-F.
Pros: Simple ranging method, typically available on
communication devices.
B. Pros and cons of location-dependent measurements for Cons: Error-prone, typically requires an extensive tuning of
mmWave localization the path loss model. RSS measurements are often affected by
Angles of arrival and departure, angle difference-of- the losses in the front-end receiver architecture of the client
arrival [121]–[127] — The term angle of arrival (AoA) refers and by the number of quantization bits in its ADC circuitry.
to the angle at which radio signals illuminate the antenna
array at the receiver. The transmitter-based counterpart, the Time information [140] — Time information is another
AoD, refers to the angle at which the radio signals emanate common proxy for the distance between two devices. Typical
from the antenna array at the transmitter front-end in order measurements used for this purpose involve ToF and TDoA
to reach the receiver. In most cases, more than one antenna measurements. ToF (also known as ToA) measurements
elements are required to compute angle information. Other exploit the time taken for a signal to propagate from the
methods to extract AoA information from the receiver array AP to the client in order to estimate the distance between
involve the use of CSI, beamforming methods, or subspace them. The client intuitively lies on the circumference of the
approaches such as the well-known MUSIC [115] and circle with the AP as the center and the distance estimate as
ESPRIT [116] algorithms. We cover angle-based approaches the radius. Multilateration methods can be used to estimate
in Section V-D. the location of the client. It is important to note that ToF
Pros: Relatively accessible information in mmWave systems, measurements require a tight synchronization between the AP
thanks to the large number of antennas in transmitter and and the client. mmWave signals offer better ToF estimation
receiver arrays. accuracy (thus better ranging resolution), owing to the large
Cons: If not associated to some range information, can bandwidth available, especially in the unlicensed bands. We
only yield location estimates in a relative coordinate system. cover time-based approaches in Section V-F.
Multipath propagation can distort angle estimates, if not Pros: ToF information is usually accurate when directly
properly modeled or compensated for. extracted from a device’s physical layer, which helps accurate
localization. Such protocols as the fine time measurement
Channel state information (CSI) [128]–[130] — CSI refers (FTM) protocol, when available on a device, can provide
to the measurable properties of a received mmWave signal very accurate timing estimates.
that relate to the propagation channel linking two devices, Cons: Requires sub-nanosecond sampling times in a device’s
e.g., the AP and the client. Different mmWave hardware ADC in order to yield a sufficiently fine range resolution.
may provide different forms of CSI. For example, patching
TP-Link’s Talon routers [131] with special firmware makes Hybrid approaches [141], [143]–[156] — Several solutions
it possible to extract receiver-side CSI in the form of one propose to fuse information from multiple sources in order
complex gain coefficient per receiving antenna, expressing to improve localization accuracy. For example, several works
the attenuation and phase shift that affect the strongest merge AoA and RSSI, or AoA and ToF estimates. We cover
 12

Fingerprints

CSI

AoA/ToF
estimation ML/DL algorithms

CIR ToF estimation

Radio map Geometric

Measurement of Signal User Location
fingerprints algorithms
Parameters
RSSI/SNR

Range
estimation

Geometric/ML/DL
AoA
algorithms

Hybrid

Fig. 9. General flow chart of the steps of a mmWave localization algorithm from the surveyed literature.

TABLE II
V ISUAL REPRESENTATION OF THE DISTRIBUTION OF RESEARCH EFFORTS FOR DEVICE - BASED MM WAVE LOCALIZATION .
G REEN ICONS REPRESENT RECENT PAPERS THAT EMPLOY SOME FORM OF MACHINE LEARNING .

I NDOOR MM WAVE LOCALIZATION

Traditional methods Tailored methods
Time
RSSI and SNR Angle information CSI-based Hybrid approaches
information
(e.g. [128],
(e.g. [133], [136]) (e.g. [140]) (e.g. [121], [123]) (e.g. [141], [142])
[129])
Client-
centric

AP-centric

AP-client
cooperation

hybrid approaches in Section V-G. cation of the devices;

Pros: Hybrid schemes usually achieve better accuracy. In 2) General algorithms that apply well-known range-based or
some purely angle-based algorithms, side information such range-free localization approaches to mmWave commu-
as RSSI and ToA can help resolve geometric translation, nications.
rotation, and scaling ambiguities. The algorithms in the first category are mainly angle-based or
Cons: The algorithms become more complex, and rely CSI-based: they infer the angle of arrival structure by leverag-
on the estimation of multiple quantities. In ill cases, errors ing, e.g., sector measurements in communication protocols.
compound and may make the location system more inaccurate Then, they use angle information to localize a device. By
than non-hybrid ones. way of contrast, the algorithms in the second category are
not necessarily mmWave-specific. These works can be further
According to our survey of the literature on mmWave lo- subdivided by considering where the algorithm mainly runs:
calization algorithms and to the above discussion, we identify 1) In client-centric algorithms, the intelligence mainly re-
two broad categories in the available literature: sides on the client, which may collect location-dependent
1) Algorithms tailored to mmWave communication proto- measurements by receiving signals from one or multiple
cols and schemes, that exploit protocol operations to APs, and by estimating its own location locally. This
extract geometric scenario information and infer the lo- approach is useful for systems that need to scale to up
 13

than client-centric approaches when the number of clients

AP-3 increases. Literature surveyed: [124], [128]–[130], [137]–
[139], [143], [144], [147], [149], [150], [152]
3) Schemes based on AP-client cooperation are based on
a shared intelligence, where both one or more APs and
the client run portions of the localization algorithm, and
AP-1 possibly exchange information to finally estimate the
Client
client location. Literature surveyed: [133], [140], [145],
[146], [148], [151]
In our scan of the literature, we observed a comparatively
AP-2
small number of works that employ a form of machine learning
to compute location estimates. We believe this is due partly to
(a) Trilateration
localization being a somewhat understood problem (whereby
the community prefers the use of understandable and optimiz-
AP-3 able signal processing algorithms rather than training black-
box machine learning models) and partly to the sometimes
daunting collection of training data. Yet, these prove a feasible
solution in some cases, e.g., when a huge database of different
location-dependent features is available, and the complexity
AP-1 of the considered indoor environment prevents straightforward
Client modeling.
Table II summarizes the above preliminary subdivision
pictorially, and conveys in what category most of the re-
search efforts has concentrated so far. We observe that a few
AP-2 approaches have considered baseline RSSI, SNR and time
(b) Triangulation measurements to localize mmWave devices. However, most
of the research moved to exploit the fine angle resolution that
large mmWave antenna arrays enable. A significant number
of works also consider hybrid approaches, which mix good
angle resolution with the extra information yielded by time- or
AP-3 RSSI-based measurements, and thus achieve greater accuracy.
Finally, we observe that a few recent works (from 2017 to the
C2
time of writing) rely on ML techniques, typically to process
C1
RSSI and SNR measurements and predict the location of a
AP-1 device. We highlight these works in green in Table II, in
AP-2
order to emphasize the emergence of this paradigm, previously
unobserved in indoor mmWave localization.
In client-centric algorithms, the client collects signal mea-
surements thanks to the interaction with different APs. The
(c) Angle-difference of arrival client then trains an ML model and employs it to estimate
its own location. For example, in [136], the client collects
Fig. 10. Illustration of the (a) trilateration, (b) triangulation, and (c) angle- SNR information to train ML regression models. In [157],
difference of arrival processes using ToF, AoA, and ADoA localization
geometries, respectively. Note that di and τi respectively denote the distance instead, the client resorts to AoA information to train shallow
and propagation delay between AP i and the client, c is the speed of light in neural networks and estimate its coordinates of the client.
air, αi denotes the AoA of the signal from AP i, and βi is the ADoA, i.e., Other works in this survey that employ client-centric machine
the difference of the AoAs from APs i and i + 1.
learning algorithms are [134] and [141].
AP-centric algorithms rely on APs collecting location-
dependent signal features that relate to the location of each
a large number of devices, as each device runs the al- client in a given environment. These radio fingerprints are
gorithm independently. Literature surveyed: [121]–[123], then used to train models to localize the client. For example,
[125]–[127], [132], [134]–[136], [141], [142], [153]– in [137], [138], the APs use the spatial beam SNR measure-
[155] ments collected during the beam training process in order to
2) In AP-centric algorithm, the intelligence resides in a create a radio map of the environment. DL models are then
computing entity connected to one or multiple APs, which trained to estimate the location and orientation of the client
coalesce their measurements from multiple clients in or- devices. Other works in this survey that employ AP-centric
der to estimate the location of each client. These schemes machine learning algorithms are [139] and [124].
are ideal for seamless network management purposes Other algorithms rely on some form of AP-client cooper-
(e.g., to optimize client-AP associations) but scale less ation to collect location-dependent signal features and train
 14

machine learning models. In these schemes, the features can for high mobility scenarios. Moreover, Polese et al. [165] pro-
be collected either by the APs and the client separately and pose a 60-GHz SDR, fully digital experimentation platform.
then exchanged, or through possibly multi-step procedures re- It uses a Xilinx KC705 and has 4 independent streams.
quiring AP-client cooperation. The only work in the literature Alternatively, commercial-grade equipment can be lever-
that uses this technique for ML models is [133]. Here, RSSI aged for localization purposes, usually by substituting the
and beam indices obtained both at the client and at the APs provided operating system image with a custom build that
after the beam alignment process are used to generate radio embeds application program interfaces (APIs) to access the
fingerprints at different client locations. output of the beam training procedure. For example, the work
Notably, Table II clearly shows that ML-based algorithms in [143] realizes a geometric 3D localization system using a
are mostly AP-centric or hinge on a cooperation between 4×8 phased array within a router that embeds a Qualcomm
APs and clients. The main reason is most that APs are QCA9006 tri-band chipset for AoA and ToF measurements.
infrastructured devices, and have easier access to compute The work in [144], instead, taps into the output made available
power in local servers through fast cabled connections. by the Talon AD7200 [131] routers’ firmware. In the latter
case, the hardware and the interface require significant adap-
tations of the angle estimation algorithms. For example, the
C. Evaluation tools for mmWave localization firmware and operating system used in [144] returned coarsely
We now look into the tools that have been used so far quantized power measurements for each beam pattern and
to evaluate mmWave localization algorithms. From the sometimes incomplete measurement outputs, which required
surveyed literature, we observe both experimentation-based to re-cast the angle estimation algorithm to be robust against
and simulation-based performance evaluation, depending on quantization noise and missing values. The proprietary setup
whether a proposed scheme is evaluated using mmWave used in [143] returns the raw CIR measurements, which are
hardware- or software-based setups. then sanitised to extract the azimuth and elevation angles
of arrival from the LoS paths, and the ToF information for
1) Experimentation-based performance evaluation: Local- distance estimation.
ization experiments so far have been carried out using either Other works such as [167] also employ COTS devices like
laboratory-grade or commercial-grade equipment. Laboratory- the 802.11ad-enabled Airfide AP [168] to enhance the antenna
grade equipment typically includes software-defined radios array performance for omni-directional coverage and to im-
(SDRs) for signal generation and a mmWave up-converter, prove link resilience in mobile and dynamic environments.
with a directional antenna to drive signal emission. For exam- Table III summarizes the above discussion by relating the
ple, the above setup is used in [127], where the authors employ works in our survey with the hardware platforms used to
horn antennas to emulate narrow beam patterns. A similar validate mmWave localization algorithms. We observe that
setup is part of the work in [158] and [140], where the authors software-defined platforms are still preferred, due to their
employ the Zynq 7045-based SDR and the universal software greater versatility and to the availability of multiple dig-
radio peripheral (USRP) X310-based SDR, respectively, in ital receiver chains. COTS hardware is starting to appear
addition to a 60-GHz analog front-end to emit the mmWave in experimental evaluations, although this typically requires
signals. The authors of [159], [160] have used an NI SDR system management (and sometimes hacking) skills to flash
with a 60-GHz transceiver that enables the user to fully the hardware with firmware and custom operating systems that
program of the physical layer (PHY), MAC, and network give access to information from the radio receiver chain.
layers, especially for wireless LAN (WLAN) applications. It From a practical standpoint, the manufacturers of
also incorporates a 24-element Sibeam reconfigurable antenna commercial-grade mmWave devices typically define a code-
array. A field-programmable gate array (FPGA)-based setup is book of antenna weights that drive beam patterns to cover
discussed in [161], where the authors have used the XCKU040 the largest set of lookout directions. As a result, the corre-
Kintex UltraScale FPGA for the baseband processing of a 60- sponding beam patterns are not necessarily narrow, nor do
GHz reconfigurable phased antenna array. PEM-003 60 GHz they necessarily present a single direction where the gain is
transceivers were used as the RF front-end for the experi- maximum [169].
mentation. Recently, the New York Uuniversity spin-off Pi- Yet, standard-compliant beam training procedures still help
Radio [162] developed dedicated SDR boards for mmWave retrieve location-dependent measurements through an auto-
wireless communications. The Pi-Radio v1 SDRs consists of mated process that is typically implemented in every device.
a 4-channel fully-digital transceiver board with a Xilinx’s For example, the 802.11ad standard [15] presents a two-phase
ZCU111 RF system-on-chip (SoC) [163], and operates over beam training process:
a bandwidth of about 2 GHz in the 57-64 GHz band. • Sector-level sweep (SLS): During this phase, the transmit-
Other platforms currently in use in experimental work ter (or beamformer) periodically transmits sector sweep
include the open source mmWave experimentation platform (SSW) frames using the different beam patterns defined
proposed in [164]. It consists of a Xilinx Kintex Ultrascale in the sector codebook. The receiver (or beamformee),
FPGA with a 60-GHz front-end. The FPGA is integrated on receives these frames omnidirectionally and sends back
an AMC599 board that implements hardware signal processing an acknowledgment with the transmit sector yielding the
and storage for real-time frame processing. It can also provide highest signal quality. Subsequently, the two devices swap
antenna array reconfigurability for fast beam switching, e.g., roles, and the receiver selects its best transmit sector. This
 15

TABLE III
S UMMARY OF THE HARDWARE AND SOFTWARE PLATFORMS USED IN MM WAVE LOCALIZATION ALGORITHMS

Hardware Platform Related Literature

Vubiq 60 GHz development system [121], [122], [127], [144], [148]

Zynq 7045 based SDR with 60 GHz analog front-end [140]
4×8 phased array AP with QCA9006 triband chipset [143]
TP-Link Talon AD7200 [128], [133], [136], [137], [138], [139]
QCA6320 baseband module with QCA6310 RF front-end [152]
USRP X310 and TwinRX daughterboard with 60 GHz analog front-end [146]
MicroTik wAP 60G [153]
Software Platform Related Literature

NYURay Ray tracer [145], [166]

S 5GCHANNEL simulator [155]

Beacon Interval (BI) Transmits SSW packets using Listens omni-directionally and
Beacon Header Interval (BHI) Data Transmission Interval (DTI) coarse sector patterns sends ACK with best sector pattern

BTI A-BFT
A-BFT ATI CBAP
CBAP SP CBAP SP

200
Fig. 11. Beacon Interval frame of the IEEE 802.11ad standard [15]. It is
important to note that after beam training process, STAs contend for the
channel during the contention based access period (CBAP) and access it
contention-free during the service period (SP).
Transmitter Receiver

(a) Sector Level Sweep

phase provides coarse-grained beam patterns that are best
suited for the two communicating devices.
• Beam refinement protocol (BRP): This optional phase Transmitter and Receiver iteratively fine-tune their beams
can be used to refine the beam patterns chosen after to achieve the best communication link

the SLS phase. The BRP process is iterative. The two

devices exchange special BRP packets requesting and
acknowledging the transmit (TX) and receive (RX) train-
ing requests (TX-TRN and RX-TRN). The result is fine-
grained beam patterns for the transmission and reception
of the data, resulting in not just better directivity and
therefore higher-throughput links, but also in a higher Transmitter Receiver

correlation between the beam pattern used and the AoA (b) Beam Refinement Protocol
of a signal.
These phases occur during the association beamforming train- Fig. 12. A simple illustration of the sector level sweep and beam refinement
ing (A-BFT) subinterval of the beacon interval (BI), as part protocol as proposed in the IEEE 802.11ad standard [15].
of the channel access mechanism. The beamforming process
during the data transmission interval (DTI) is to handle device
mobility, blockage, etc. The BI frame for channel access When run with generic beam patterns, the above procedures
is shown in Fig. 11 and the two beamforming phases are do not yield a one-to-one relationship between the angle of
illustrated in Fig. 12. arrival or departure of a mmWave signal and the antenna
The more recent 802.11ay standard [170] formalized beam configuration that leads to the highest received power. Yet,
training procedures that enhance those of 802.11ad, namely if the beam patterns of the codebook are known, a mmWave
the beam refinement protocol transmit sweep (BRP TXSS) device can still estimate angles of arrival via signal processing
and the asymmetric beamforming training (ABT) [17]. These techniques involving compressive sensing [169], or linear
procedures rely on a channel reciprocity assumption to speed programming and Fourier analysis [144]. Knowing angles
up beam training (through the BRP TXSS scheme) and slightly of arrival enables angle-based localization techniques, with
improve the process to compensate for the possibly different the additional advantage that angle estimation hinges on
antenna gains at the AP and at the client. standard beam training procedures, with no need for external
In addition, 802.11ay speeds up training in the presence hardware components. In other words, localization becomes
of several clients through group beamforming, which extends an embedded feature of mmWave communications.
beam training to manage multiple clients simultaneously.
 16

2) Simulation-based performance evaluation: Simulation The simplicity of these algorithms motivated the authors
is the performance evaluation tool of choice if mmWave hard- of [121] to generalize the schemes in [122] for any number
ware is not available or if the available platforms do not offer of APs. These algorithms are extensively simulated as well
sufficient flexibility to measure location-dependent features. as experimentally validated on 60 GHz COTS devices, in
A common practice observed in the literature is to employ different indoor scenarios against two benchmark algorithms
ray tracers to mimic the propagation of mmWave signals. based on fingerprinting and AoA. The two algorithms provide
These ray tracing simulators are typically designed based on sub-meter accuracy in most indoor environments with multiple
the channel models described in Section II. The main idea is antennas. Triangulation-based scheme performs slightly better
to simulate the mmWave wireless channel characteristics at than the ADoA-based one in most scenarios, but independence
various indoor locations. Besides allowing the experimenter of orientation and compass bias makes ADoA more preferable.
to measure channel features, ray tracers help create a radio The ideas proposed by [122] have also been used by the
map of the environment, and can thus substitute costly and authors of [125] for context inference and obstacle detection.
time-consuming measurement campaigns [171]. They use the TV and ADoA algorithms for receiver localiza-
Two examples of such simulators are NYURay, a 3D tion using one AP, estimate the locations of virtual anchor
mmWave ray tracer developed by New York University [171], nodes, and thus infer the presence of obstacles.
and S 5GChannel, developed by Siradel. NYURay was ini- AoA measurements have also been used for simultaneous
tially conceived as a geometry-based 2D ray tracer and was localization and mapping (SLAM). For example, in [126],
used in [166] to investigate indoor positioning algorithms the authors propose a joint access point and device localiza-
based on AoA, combined path-loss and AoA, or RSSI val- tion (JADE) algorithm that jointly maps the location of the
ues. NYURay was later extended in [145] to support 3D client and of the physical and virtual APs, while mapping
ray tracing by combining the shooting-and-bouncing rays the indoor environment, without any prior information (i.e.
(SBR) technique [172] and the geometry-based technique. number of access points, boundaries of the room, etc.). The
NYURay found extensive use, not just in indoor environ- algorithm measures AoAs from the beam training procedure
ments, but also outdoors [173], [174]. Siradel developed the and leverages ADoAs to estimate the location of the APs and
S 5GChannel [175] 5G channel simulator to address the then of the client. Environment mapping follows by matching
challenges of 5G signal propagation at mmWave frequencies physical and virtual anchors and by predicting reflection
indoors and outdoors. S 5GChannel’s ray model has been used points on surrounding surfaces. Simulation results show sub-
in [155] to develop a framework for joint localization and meter accuracy in 90% of the cases, even for erroneous AoA
mapping. estimates. JADE outperforms the approaches in [121] in almost
A few additional works in the literature evaluate their all scenarios.
proposed schemes using custom simulation software typically A similar algorithm that exploits AoA information to derive
written in MATLAB or Python. The general purpose of ADoA estimates and fuses multiple measurements at different
such software is to generate synthetic datasets with realistic locations is CLAM [127]. Like in [126], the algorithm pro-
mmWave propagation characteristics, although typically re- ceeds by first estimating the location of the anchor APs, then
stricted to the specific signal properties required for each study of the client, and finally of the environment’s boundaries. The
(e.g., AoA values, ToA measurements, etc.). algorithm is simulated and experimentally evaluated, showing
In the following subsections, we explain the details of each sub-meter device localization errors in about 90% of the cases.
surveyed work, and provide a synopsis of the main results of A recent work explores deep learning-based localization
each paper and of the main enabling techniques in the form scheme. The authors of [157] propose a shallow neural net-
of summary tables at the end of the section. work model to estimate the coordinates of the client device
in an indoor environment, using ADoA measurements. The
D. Angle-based algorithms network is trained with imperfect location estimates from the
AoA measurements, alongside the quasi-optical nature of JADE algorithm [126], which jointly estimates the location
mmWave signal propagation, facilitate high-accuracy localiza- of the APs and the clients with zero knowledge of the envi-
tion based on triangulation. This is the simplest approach to ronment. This relieves the burden of explicitly collecting the
localization using AoA, wherein the angle information from training dataset. The performance evaluation of the proposed
the transmitting APs and simple geometric principles are used scheme results in sub-meter client localization accuracy in
to compute the client’s position. In a 2-D plane, such position ≈ 90% of the scenarios, even with large AoA errors.
can be estimated using just two APs [176]. In [123], the authors present mobile device positioning
Geometric methods are the simplest methods for localiza- scheme in an indoor mmWave massive multiple-input single-
tion when using AoA estimates. In [122], the authors present output (MISO) scenario. The two-fold scheme utilizes coarse-
three lightweight single-anchor algorithms based on the AoA grained AoD information from mobile clients with a single
measurements. These algorithms are based on triangulation, antenna to estimate the position of each client via downlink
ADoA, and fingerprinting, respectively. The algorithms have transmissions using adaptive beamforming.
been simulated and also experimentally validated on pre- We can observe that angle-based algorithms usually rely on
standard mmWave hardware operating at 60 GHz, showing geometric approaches for device localization. However, ML
that they achieve sub-meter accuracy with high probability, and neural network regression models can also be used to
given the AoA estimate errors are low. learn a non-linear mapping between AoA measurements and
 17

client locations. devices. Probabilistic location models are generated based on

the fingerprint data and are leveraged for location estimation.
E. Channel information-based algorithms The algorithm is experimentally evaluated using 60 GHz
COTS devices. Many times, SNR-based fingerprinting is also
In recent works, mmWave CSI is also used to estimate the
at the core of some mmWave localization works, especially
location of the client. The definition of CSI varies from work
in combination with machine learning and deep learning
to work. Typically, the term refers to the complex amplitude
techniques. The authors of [21], [136] propose machine learn-
of the channel gain perceived at a receiving antenna, or to the
ing regression models for localization in warehouses. SNR
vector of such gains measured by all elements of an antenna
information is collected from Talon AD7200 routers. The
array. A work exploiting CSI for localization is [129], where
supervised regression models are trained offline and then
the authors present a channel parameter estimation method
deployed for localization at run time. The proposed method
that transforms the mmWave uplink training signal into a
achieves sub-meter accuracy in 90% of the cases.
higher-dimensional tensor using the canonical polyadic model.
Similar machine learning regression models have been used
Tensor factorization using the proposed generalized structured
for location estimation in [137], where the authors use spatial
canonical polyadic decomposition results in time delay, AoA,
beam SNR values, typically available during the beam train-
and path fading coefficient estimates. These parameters are
ing phase, in order to generate a location- and orientation-
used to localize and track a mobile device.
dependent fingerprint database. Deep learning techniques are
A different way to exploit uplink CSI estimates [128]
also the main enablers for localization in [138] and [139],
requires that the APs convert the LoS CSI measurements into
where the authors proposed ResNet-inspired models [177] for
angle information and then localize the client. The system is
device localization in LoS and NLoS scenarios. To tackle
implemented on Talon AD7200 routers (without interfering
the challenges imposed by NLoS conditions, the authors use
with 802.11ad operations), and the authors propose to employ
spatial beam SNR values in [138], whereas they employ multi-
the location estimates to optimize AP–client associations. The
channel beam covariance matrix images in [139].
system achieves sub-meter localization accuracy in about 80%
One example of how ToF measurements have been used in
of the cases.
the mmWave context is presented in [140]. Here, the authors
With a focus on localizing passive objects, in [130] the
present a two-way ranging based on round-trip ToF (RTToF)
authors use the CIR captured after reflection from different
information. The scheme estimates the distance between mas-
objects and surfaces in an indoor environment to detect
ter and slave nodes, and then trilaterates the position of the
objects and also model the indoor environment in 2D. The
slaves. The authors implement their algorithm on an SDR with
proposed method has been evaluated using a testbed developed
a 60 GHz SoC. The proposed system achieves an average
specifically for this purpose.
distance estimation of 3 cm and an average positioning error
The use of CSI for localization is comparatively new
below 5 cm.
for mmWave indoor device-based localization, most likely
Although conventional wireless localization schemes rely-
because retrieving full CSI or CIR data requires low-level
ing on RSSI or SNR measurements employed trilateration,
hardware access, and only a few experimental firmware ver-
machine learning-based fingerprinting algorithms are gaining
sions provide it. However, CSI and CIR can map to angle
more popularity for mmWave-based localization systems. This
and time information, and therefore represent a promising
is due to the availability of mid-grained channel measure-
and practical research direction, especially as feature-richer
ments from the beam training procedures of 5G and IEEE
mmWave hardware and firmware emerges.
802.11ad/ay systems [139]. These techniques provide higher-
accuracy location estimates compared to conventional tech-
F. RSSI and ToF niques.
RSSI and SNR based localization systems generally employ We also observe that mmWave systems do not rely on purely
trilateration or fingerprinting-based techniques to localize the time-based measurements for localization. Even though the
client. A number of works in the literature illustrate this large bandwidth of mmWave signals can provide fine time
concept. The authors in [132] investigate trilateration-based measurements, such measurements tend to be fully available
localization algorithm using RSSI measurements for 60-GHz only on custom high-end mmWave transceivers. Therefore,
IEEE 802.11ad WLANs. They modify the trilateration algo- many schemes tend to collect other signal measurements as
rithm based on the concept of (weighted) center of mass. well.
Simulations on randomly generated data points and the RSSI
measured based on the IEEE 802.11ad channel model result G. Hybrid approaches
in an average positioning error of about 1 m. This is among A combination of two or more techniques mentioned above
the earliest works on mmWave-based indoor localization that can be used to build systems that achieve better localization
leverages RSSI measurements. or mapping accuracy, with respect to stand-alone techniques.
RSSI is also the foundation of several fingerprinting-based Coupling different sources of information is useful in chal-
localization schemes, especially in sub-6 GHz wireless net- lenging environments, where some mmWave parameter mea-
works. The authors of [133] propose a localization system surements may fail.
that generates fingerprints of transmit beam indices and the Angle information along with RSSI-based ranging are the
corresponding RSS measurements between a pair of mmWave foundation of several mmWave localization approaches in the
 18

literature. The authors in [147] propose a positioning algorithm level distance estimation and decimeter level 3D localization
using RSS and AoA measurements. These measurements are accuracy (median error 75 cm and sub-meter error in 73% of
derived from a channel compression scheme designed for the cases) in a realistic indoor environment. The system has
a mmWave mMIMO scenario with only one AP. The RSS also been evaluated in various experimental conditions.
and AoA estimates from the above methods are employed The author of [145] propose a map-assisted positioning
for position estimation. The system provides decimeter-level technique using the fusion of ToF and AoD/AoA information.
accuracy even at low SNR, and even lower errors as the SNR A 3D map of the environment is either generated on the fly or
increases. assumed to be known a-priori. The scheme measures a set of
As opposed to ranging, the algorithms proposed in [141] possible user locations by fusing the estimated ToF values with
and [134] are based on location fingerprinting. In particular, angle information. These estimated locations are clustered, and
the authors measure RSSI and AoA information at various the cluster centroid is output the final location estimate. The
reference points in an indoor environment to generate location algorithm is simulated on the data collected at 28 GHz and
fingerprints. K fingerprints nearest to the client measure- 73 GHz by NYURay 3D ray tracer. The best-case and the
ment are selected from the dataset, and the location estimate worst-case mean localization error is found to be about 12 cm
corresponds to the weighted average of these K reference and 39 cm respectively.
points. The algorithm has been simulated with 2.4 GHz and Instead of explicitly fusing ToF and AoA information,
60 GHz, showing that the average position error is much lower the authors of [148] propose a pseudo-lateration protocol,
for mmWave signals than lower-frequency signals. To solve that enacts the three following steps: i) sector sweeping for
the problem of collecting a sufficiently large dataset, [134] tracking LoS and NLoS paths to compute physical and virtual
generates 3D beam fingerprints using RSSI and beam infor- anchors, respectively; ii) angular offsets measurements using
mation. Weighted K-NN was used to localize an unmanned extended sector sweeping; and iii) ToF measurements for
aerial vehicle (UAV) in GPS-denied indoor environments. distance estimation. A post-processing stage is employed for
Particle filters were used along with the imperfect location position estimation. The protocol has been simulated and
estimates to track the motion UAVs. The proposed scheme was implemented using a 60 GHz mmWave testbed. The protocol
experimentally validated, and the results showed sub-meter implementation achieves centimeter-level location estimation
positioning accuracy on average. accuracy within 1.5 m and decimeter accuracy beyond 1.5 m.
RSS jointly with AoA information enables mmRanger [152] The authors of [149] explored adaptive filters for motion-
to autonomously map an indoor environment without infras- assisted indoor positioning. An improved LMS filter estimates
tructure support. The mmRanger scheme senses the environ- the AoA of the client by using the client location, velocity
ment and uses time domain RSS sequences to reconstruct the and measured ToF as the inputs. AoA and ToA estimates
path geometry via a path disentanglement algorithm. Then, are fed to an unscented Kalman filter (UKF) to track the
AoA and RSS information from the reflecting surfaces are client’s position. The two-stage algorithm is simulated in an
exploited to reconstruct the geometries of each surface. More- office environment with one AP and achieves centimeter-level
over, a robot pedometer assembles all estimated fragments to positioning accuracy.
form a complete map of the environment. The results of the Because mistaking LoS for NLoS paths may offset location
proposed system implementation show a mean estimation error estimates significantly, the authors of [150] propose a scheme
of 16 cm for reflection points, and a maximum error of 1.72 m. to tell apart mmWave LoS and NLoS MPCs having incurred
In [144], the authors leverage coarse-grained per-beam up to one reflection. For this, they use TDoA and AoA
pattern SNR measurements provided by a modified operating information and apply the mean shift clustering technique.
system flashed on multiple TP-Link Talon AD7200 802.11ad- Then, they apply an AoA-based localization scheme that
compliant COTS mmWave router. The AoA estimation prob- computes least-squares estimates. The methods show a 98.87%
lem is formulated using linear programming, and the location accuracy in path identification and positioning error of less
is estimated using a modified particle filter and a Fourier than 75 cm in 90% of the cases. NLoS scenarios have also been
analysis-based goodness function. The proposed scheme is exploited in [151], where the authors propose a positioning
experimentally validated and the system achieves sub-meter scheme that relies on differential angle information, which
accuracy in 70% of the cases. AoD and SNR information is independent of angular reference. This scheme has been
were used in [142] to design beam-based midline intersection evaluated in an indoor environment with a geometric ray tracer
and beam scaling-based positioning algorithms. These were based on an IEEE 802.11ay channel model, and achieves sub-
evaluated using both ray-tracing simulation and a WiGig SoC 30 cm position estimation errors in 90% of the cases.
transceiver. The experiments, carried out under LoS condi- In [154], the authors present schemes for localization,
tions, yielded centimeter-level location estimation errors. mapping, obstacle detection and classification. Localization
Time-based measurements are often enriched with angle and mapping make use of AoA and ToA measurements to
information in order to achieve better positioning accuracy, estimate the location of the receiver and of virtual anchors.
especially for mmWave systems. For example, in [143], the The latter are used to detect obstacles by estimating reflection
authors propose mWaveLoc. The proposed system uses mea- points. Snell’s law and the relationship between the RSS and
sured CIRs to calculate AoA and ToF data. The system is the reflection coefficient are used to classify the obstacles
implemented on IEEE 802.11ad off-the-shelf devices leverag- based on material composition. The presented algorithms have
ing the OpenWRT operating system, and achieves centimeter- been simulated in an indoor environment.
 19

Besides locating a client, the schemes presented in [154] Similar issues affect the estimation of time information
have been integrated into a SLAM framework in [155]. through CIR or packet exchange means. We can obtain fine
This framework involves algorithms for localization, obstacle time measurements thanks to the large bandwidth of the
mapping and tracking. Extended Kalman filter (EKF)-based mmWave signals. However, this requires a very tight synchro-
tracking helps improve obstacle detection and mapping. The nization between the AP and the client devices.
framework has been simulated in an indoor environment, ML-based algorithms – Owing to recent contributions, we
yielding sub-meter errors in 90% of the cases. In the same observe a paradigm shift towards self-learning location sys-
context, the EKF improves the obstacle mapping accuracy to tems that exploit the information from mmWave signals.
sub-centimeter. In these works, ML and DL-based models use information
In [146], the authors present a device localization scheme, extracted from mmWave signals at different locations to form
where the AP and the client are equipped both with sub- a dataset and eventually learn an accurate model to estimate
6 GHz and with mmWave technology. Sub-6 GHz antennas client positions. However, most of these models are specific
are used for AoA estimation and mmWave antennas are fed for the location the training data comes from, and do not
with the AoA estimates for subsequent beam training and two- translate well to other locations. Most of the algorithms in
way ranging. The proposed method has been experimentally the literature train ML and DL models through RSSI or SNR
validated using SDR platforms, both in an anechoic chamber fingerprint maps. Recent works have showed how AoA and
and in an office environment. Results show 2◦ AoA errors and CSI information also help ML models learn a non-linear
centimeter-level ranging accuracy in the anechoic chamber, function either to estimate the location of the client or to
and 5◦ AoA error with an average 16-cm range error in the associate a client to the best APs.
second one. Although these systems provide good localization accuracy,
In [153], the authors propose to track the changes in the they also face several challenges: the collection of large
CIR measured at the station, that is equipped with an FPGA- training dataset; the computational complexity which limits
based platform with IEEE 802.11ad, in order to localize a their application to COTS or embedded devices; their depen-
device-free object in an indoor industrial environment. The dence on the training environment. ML methods have thus
station uses the estimated CIR to measure the AoD and ToF found comparatively limited application to date. A valuable
of the signal reflecting off a moving object. Tracking CIR contribution to the community would be a collaborative effort
changes over time helps classify the reflections as static or towards a public benchmark dataset, that different authors
mobile. Then, a Kalman filter smooths the trajectory of the would use to feed different machine learning approaches.
mobile object. The results show sub-meter location errors in Error mitigation in mmWave localization – Errors in signal
all scenarios, and a mean accuracy of 6.5 cm. measurements due to imperfect signal parameter estimation
From the literature surveyed above, we can observe that limit the performance of localization systems [30]. These
most localization schemes use angle information along with errors are often due to the unpredictable interference between
RSSI/time information, and often rely on geometric algorithms multiple propagation paths and the fading that results, or to
to compute high-accuracy location estimates. The use of NLoS arrivals reaching a device [178]. A detailed mathemati-
adaptive filters such as least mean squares (LMS) and Kalman cal analysis for error mitigation is presented in [30]. In the case
filters helps mitigate location estimate errors, especially with of mmWaves, measurement errors may affect angle and time
mobile clients. measurements. The works in the literature resort to adaptive
filter-based techniques mostly to mitigate the location esti-
H. Summary, highlights, and challenges mation errors resulting from localization algorithms fed with
We now summarize the surveyed literature in order to high- error-prone data [19]. For example, the approaches in [134],
light the main pitfalls and lessons learned from the methods. [144] resort to particle filters to mitigate client location errors.
Geometry-based algorithms – The algorithms based on Different types of Kalman filters are another typical choice
geometric techniques mostly rely on angle information to smooth out client location estimates and trajectories [179].
(AoA/AoD) for localization. As mmWave signals propagate The authors of [150] use LMS filters to mitigate large errors in
quasi-optically, angle information becomes a reliable means to AoA and ToF measurements. A detailed account of the tools
estimate the direction of the source. RSSI and ToF information and techniques employed in each surveyed paper is provided
help estimate the distance between a mmWave source and its in Table VI at the end of this section.
receiver; thus, applying geometric methods such as triangu- Table IV supports the above discussion by summarizing the
lation and trilateration can help localize a client. However, techniques employed in each of the surveyed works. We ob-
the accuracy of such algorithms depends upon the accuracy serve a preference for geometry-based localization approaches,
of angle and time measurements, and most of them require with different supporting signal processing algorithms.
accurate indoor floor plan information to work reliably.
From the perspective of COTS devices, angle measurements I. Discussion and future research directions
are obtained either by decomposing CSI measurements using We summarize the findings of our survey in Table VI.
parameter estimation techniques or from the beam patterns The table succinctly conveys the main proposition of each
chosen after beam training. However, imperfect beam patterns paper, the main techniques used among those outlined in
with broad main lobes and non-negligible sidelobes can lead Sections V-D to V-G, the tools employed, and a high-
to angle estimation errors. light of the performance attained. We observe a number of
 20

TABLE IV
S UMMARY OF THE MAIN TECHNIQUES USED IN THE SURVEYED PAPERS

Analytical Tools Related Literature

Beamforming techniques [123]

Clustering methods [129], [150], [152]
Deep learning [124], [138], [139], [157]
Fourier analysis [144]
Geometry [121], [122], [125], [132], [135], [142], [143], [145], [154], [156]
Kalman filters [128], [149], [153], [155]
Least mean square filters [149]
Levenberg-Marquardt (LM) method [151]
Linear programming [144]
Machine learning models [124], [134], [136], [137], [141]
MUSIC [146]
Particle filters [134], [144]
Probabilistic data modelling [133]
Tensor analysis [129]

TABLE V
S UMMARY OF THE EVALUATION METHODS USED IN THE MM WAVE LOCALIZATION ALGORITHMS

Evaluation Related Literature

Experimentation [121], [122], [127], [130], [133], [134], [136], [138], [139], [140], [142], [143], [144], [146], [148], [152], [153]
Simulation [123], [125], [129], [132], [135], [141], [145], [147], [149], [150], [151], [154], [155], [156], [157]

mmWave localization approaches exploiting different features cost mmWave platforms that implement standard-compliant
of mmWave signals as well as different properties of mmWave operations while still providing APIs for researcher and de-
propagation. While some schemes rely on well-known tech- velopers to access low-level physical layer measurements,
niques, e.g., based on ToF and RSSI measurements, even such as per-beam pattern CSI, or even better, full CIRs. This
these techniques have been further developed to leverage the would democratize the research on practical algorithms that
sparsity of mmWave multipath patterns in order to collect fully integrate localization as part of standard communications
more precise measurements with a finer time resolution. In operations. In particular, such platforms would help research
some environments, typically with special-purpose lab-grade better algorithms to manage scenarios featuring multiple APs,
mmWave hardware, the corresponding localization schemes which are expected to be common in indoor mmWave de-
often yield decimeter- or sub-decimeter-level accuracy, but ployments. Moreover, there is ample space for the design of
require specific protocols to exchange the data the algorithms zero-initial knowledge algorithms that require no input data
need. from the user, and autonomously bootstrap the algorithm by
With respect to such approaches, angle-based localization finding the location of all anchors (e.g., all APs), localizing the
schemes relying on AoA and ADoA measurements still prove clients, and using the joint location information of all clients
accurate, and yield the additional benefit that AoA measure- and APs to estimate the floor plan of indoor environments
ments can directly result from beam training operations at in a SLAM fashion, both as a stand-alone solution and
link setup time. Hybrid solutions that leverage both angle as a complement to the device-free radar-based approaches
information and time/RSSI information tend to show even described in Section VI.
better performance, although only a minority of them has been From the point of view of ML schemes, we observe that
tested in operational environments with COTS equipment. most approaches still require lengthy training data collec-
Finally, Table V summarizes the performance evaluation ap- tion operations before achieving practical accuracy levels.
proach taken in each surveyed paper. We observe a majority of Moreover, a trained ML algorithm remains specific to the
evaluations based on experiments with real hardware, although area where training data was collected. Therefore, further
simulation is still used in several contexts, e.g., as a tool to research is needed on machine learning approaches that work
quickly and affordably generate large datasets. with less training data, federate training results from different
clients and APs in order to speed up the training phase, and
mmWave technologies are expected to keep gaining momen-
can be transferred across different, even previously unseen
tum as part of the 5G-and-beyond ecosystem, and there exist a
environments.
wealth of promising research directions to realize the vision of
embedding localization as a feature of mmWave communica-
tions. According to our analysis, we identify the following
key research directions. The community needs more low-
 21

TABLE VI: S UMMARY OF THE LITERATURE ON INDOOR MM WAVE LOCALIZATION

Proposition Techniques Tools Used Performance

Localization Algorithms
Round-trip time based localiza- Distance estimation error within 3 cm and
ToF Geometry
tion [140] location estimation error within 5 cm.
3-D localization with median accuracy of
Accurate 3D indoor localiza-
AoA-ToF Geometry 75 cm with sub-meter accuracy in 73% of
tion using a single AP [143]
the cases.
Improving localization accu- Linear programming, Fourier
AoD, RSSI Sub-meter accuracy in 70% of the cases.
racy [144] analysis, Particle filters
Improving the accuracy of de- Sub-meter localization accuracy in 70% of
AoA Geometry
vice localization [121], [122] the cases.
Sub-meter localization accuracy in 80%
Improving location estimation
Regression trees, Extended cases and throughput improvement be-
accuracy and network perfor- CSI
Kalman filter tween 8.5% and 57%, lesser outage prob-
mance [128]
ability, SNR within 3 dB of optimum.
Fingerprinting based indoor lo- Mean and median localization error of
RSSI Probabilistic models
calization [133] ≈30 cm.
Multi layer perceptron regres- Centimeter-level accuracy with root mean
Indoor localization for intelli-
SNR sion, Support vector regression, square error (RMSE) of 0.84 m and MAE
gent material handling [136]
Logistic regression of 0.37 m.
Avg. RMSE is 17.5 cm with coordinate
Fingerprinting based indoor lo- Machine learning algorithms, estimates within 26.9 cm in 95% of the
SNR
calization [137], [138] Deep learning cases. Median and mean RMSE of 9.5 cm
and 11.1 cm respectively.
Fingerprinting based indoor lo- Location classification accuracy greater
Deep learning, Machine learn-
calization in NLoS environ- SNR than 80%. Median location estimation error
ing algorithms
ments [139] of about 11 cm.
Map-assisted indoor localiza- AoA, AoD, Mean localization accuracy of 12.6 cm and
Geometry
tion [145] ToA 16.3 cm in LoS and NLoS respectively.
Sub-6 GHz-assisted device lo- AoA estimation error less than 5◦ and
ToF-AoA MUSIC, Geometry
calization [146] about 16 cm distance estimation error.
Single-antenna client localiza- 60% improvement in the accuracy in the
tion using downlink transmis- AoD Adaptive beamforming downlink scenario as compared to in the
sions [123] uplink scenario.
Indoor positioning for wide-
Tensor decomposition, Cluster-
band multiuser millimeter wave CSI Decimeter-level position estimation errors.
ing methods
systems [129]
Indoor network localiza-
RSSI Geometry Mean positioning error around 1 m.
tion [132]
Low complexity channel compression
3D indoor positioning for
and beamspace estimation developed.
mmWave massive MIMO AoA-RSSI Geometry
Decimeter-level positioning errors
systems [147]
achieved in NLoS scenarios.
Average positioning error for mmWave is
Location fingerprint-based lo-
AoA-RSSI K-nearest neighbours 4 times less compared to lower frequency
calization [141]
signals.
UAV positioning in GPS- Sub-meter 90th-percentile location errors
RSSI Weighted K-NN, Particle filters
denied environments [134] in different cases.
Centimeter-level estimation error in all
Beam-based UE positioning in cases. Experimental results approach the
SNR, AoD Geometry
indoor environment [141] simulations results with MSE difference of
0.1 m.
Centimeter accuracy in location estimation
Single RF chain-based local-
ToF-AoA Geometry within 1.5 m and decimeter accuracy be-
ization [148]
yond 1.5 m.
Motion feature-based 3D in-
AoA-ToA LMS and Kalman filters Centimeter-level positioning accuracy.
door positioning [149]
 22

TABLE VI: S UMMARY OF THE LITERATURE ON INDOOR MM WAVE LOCALIZATION ( CONTINUED )

Proposition Techniques Tools Used Performance

98.87% accuracy in path identification and
LoS and NLoS path identifica- Mean shift clustering, Geome-
TDOA, AOA positioning accuracy ≤ 0.753 m in 90% of
tion and localization [150] try
the cases.
Positioning accuracy within 30 cm in 90%
NLoS mmWave indoor posi- AoA, AoD, Geometry, Levenberg-
of the observations with differential angle
tioning [151] ToA Marquardt (LM) method
information along with time information.
Single-beam geometric model for indoor
5G mmWave indoor position- positioning. Mean error of 0.7 m for sta-
RSSI Geometry, Beamforming
ing [135] tionary in LoS and 2.4 m for a mobile user
in LoS/NLoS scenario.
Data-driven indoor localiza- Multi-layer perceptron, Deep
AoA Sub-meter localization accuracy.
tion [124] learning
Sub-meter localization accuracy in ≈ 90%
Indoor localization with imper-
AoA Deep learning of the cases, when trained with client lo-
fect training data [157]
cation estimates from JADE algorithm.
Localization and Mapping Algorithms
Autonomous environment Reflection point mean estimation error of
RSSI, AoA Geometry, K-means clustering
mapping [152] 16 cm with a max error of 1.72 m.
Passive object localiza- Sub-meter accuracy in all cases with
AoD, ToF Kalman filters
tion [153] 6.5 cm mean error accuracy.
Sub-meter accuracy in 70% of the cases.
Localization and obstacle de-
AoA Geometry High accuracy obstacle detection and ob-
tection [125]
stacle limits estimation.
Sub-meter localization accuracy in 90% of
Localization and
AoA Geometry the cases. SLAM without any prior knowl-
mapping [126], [127]
edge.
Accurate object detection Accuracy of about 2 cm achieved in most
CIR Geometry
[130] experiments.
Simultaneous localization Sub-meter device localization accuracy in
AoA, ToF, Geometry, Extended Kalman
and mapping without a-priori 90% of the cases. Sub-centimeter obstacle
RSSI filters
knowledge [154], [155] mapping accuracy.
3-D localization and map- RSSI, AoA, Perfectly maps the environment for AoA
Geometry
ping [156] ToA errors ≤ 5◦ .
 23

VI. MM WAVE RADAR - ENABLED DEVICE - FREE reflected signal from the surrounding environment. Through
LOCALIZATION AND SENSING some signal processing algorithms (usually, Fourier transform-
based), it is then possible to estimate the distance, angle,
A. Introduction velocity, and, to some extent, the shape of the targets. TX
In this section, we focus on mmWave-based radar systems and RX are usually co-located within the same device: the
that operate over short distances (a few tens of meters), which transmitter sends a first version of the modulated signal and the
have recently emerged as a low-power and viable technol- receiver detects its back-scattered copy from the surrounding
ogy for environment sensing. These devices are expected to environment, after a very short time delay.
find extensive use in a number of relevant applications, by Modern radar systems utilize two main wave functions;
replacing standard camera-based systems, either fully or in pulsed wave (PW) and frequency-modulated continuous wave
part. Survey papers have been recently published on mmWave (FMCW). While radars are traditionally used to detect and
sensing, with a focus on signal processing [180] (both with track objects that move in the far field, such as vehicles
traditional and machine learning-based approaches) and ap- and airplanes, here we are concerned with indoor or city
plications [181]. In the following review, we emphasize the environments where the objects to be tracked may be cars or
main signal processing algorithms that are being successfully humans. Moreover, besides the tracking, vital signs can also
exploited for indoor sensing, discussing their pros and cons. be monitored such as the breathing rate and the heart rate.
In doing this, we especially focus on neural network (NN) Although these recent applications share common features
algorithms, discussing their different flavors, and exposing with traditional (far field) approaches, they also exhibit major
the most promising directions for research and development. differences due to the short distance of the radar from the
We also comment on the level of maturity of this technol- targets (near field).
ogy, i.e., about whether the proposed techniques are robust
and work without requiring environment-specific and manual B. Pulsed Wave Radar
calibration. Our analysis will also discuss on the role of the With PW radars the electromagnetic waves from an antenna
supporting architecture, which should provide communication are emitted in short bursts. The logic behind PW is to wait
and computing/processing capabilities, and on the opportunity for the reflections from the previously transmitted signal to
of implementing networks of radar devices. This would ex- reach to the antenna before sending the next burst. Thus, the
tend current systems, which often involve a single radar, to reflected signal from the initial emitted sequence of pulses are
the large-scale monitoring of physical spaces. An illustrative sampled via a secondary sequence of pulses with a different
overview of the main techniques used for sensing applications repetition time. In PW radars, energy of the transmitted pulse
that exploit mmWave radars is provided in Fig. 13. is relatively small due to the limited peak amplitude. This
Two main components of a radar device are TX and RX limitation in amplitude together with sequential sampling
RF antennas, which are combined with an ADC, micro- limits the dynamic range and results in a relatively poor SNR
controller units (MCUs), digital signal processors (DSPs) and at larger distances. For these reasons, PW-type radars have
a clock. The main idea behind such a system is to transmit fallen out of favor, and are not used for the applications
a properly shaped radio wave (e.g., pulses or continuous that will be discussed next, which are mainly about object
waves) and estimate the modifications that occur in the back- and people detection, tracking/identification, and vital sign
scattered copy of such wave, i.e., which is returned as a monitoring. From the next section onward, we thus concentrate

DBSCAN (e.g. [182])

Probabilistic Data Matching (e.g. [192])
Kalman Filter (e.g. [183])

Geometry (e.g. [186])

Clustering (e.g. [184]) Fourier Transform (e.g. [187]) Hardware Design Optim. (e.g. [193])

Machine Learning (e.g. [188])

Non-Max Suppression (e.g. [185]) Deep Learning (e.g. [189]) Feature-based Optim. (e.g. [194])

Optimization Algorithms (e.g. [190])

Statistical Modeling (e.g. [191])

Human Activity Recognition

Health Monitoring
Object Detection

Fig. 13. Overview of techniques used for mmWave radar sensing applications.
 24

on FMCW systems, which typically are the technology of is the wavelength. Hence, the distance d to an object, the so
choice for medium and larger ranges. Still, for short range called range d = τ c/2, can be retrieved as d = φ0 c/(2πfc ) =
applications, such as gesture tracking, PW-type radars might φ0 λ/(4π). When multiple objects are present, a single TX
still be a viable alternative. chirp results in the reception of multiple reflected (RX) signal
copies. According to the different time delays (τ ) between
the TX and each of the RX chirps, multiple IF signals are
C. Frequency-Modulated Continuous Wave Radar
computed, and range measurements for each corresponding
As the name implies, FMCW radars transmit a frequency- object are derived. The range resolution dres = c/(2B), highly
modulated signal in a continuous fashion. Due to the larger depends on the bandwidth B of the radar [196]: it can be
temporal duration of continuous-wave signals, FMCW yields improved by increasing the bandwidth swept by the chirp,
a much larger energy on the emitted signal as compared to yielding a longer IF signal and, in turn, leading to a more
PW. In order to cover the desired frequency band, the signal is precise reading of the environment.
linearly modulated over time starting from the lower frequency Velocity Measurement and Resolution: In an FMCW radar
within the band to the higher frequency (or vice-versa). This the velocity computation (commonly referred to as Doppler)
type of signal is most frequently referred to as a chirp, and can be achieved using two TX chirps. Initially, the object range
the linear modulation of the signal is called frequency sweep. is calculated by applying a FFT to the RX chirps. This range
An analogue continuous-wave signal can be generated with a calculation is commonly called range-FFT. The range-FFT
voltage-controlled oscillator (VCO), providing flexible adjust- of separate chirps at the same location will yield different
ments to the sweep duration independent of the bandwidth. phases. The object velocity is then derived according to the
A frequency synthesizer together with a VCO can be used phase difference between the two chirps as v = λ∆φ/(4πTc ),
to provide a digital alternative. This technique also provides where ∆φ is the phase difference and Tc is the chirp duration.
a higher spectral purity which makes it possible to avoid However, in the case of multiple moving objects having the
accidental emission of frequencies adjacent to the desired same distance from the radar, the above method no longer
band, and thus to comply with given regulations. In FMCW works. To overcome this, the radar needs to transmit N chirps
radars, the received signal is multiplied by the TX signal. The with equal separation, i.e., a so called chirp frame. When
intermediate-frequency signal component that results is then the chirp frame is passed through the range-FFT, it yields
isolated via low-pass filtering. Additionally, a low-cost ADC a phase difference containing combined phase differences
can be used to convert the received signal into the digital of all the moving objects. The result of the range-FFT is
domain. Due to the recent developments on radar hardware, passed through a second FFT called Doppler-FFT to identify
the wider operating frequency range and the above mentioned specific phase differences ω of each object. In the case of two
advantages, FMCW radars are currently preferred over PW objects, the corresponding phase differences, ω1 and ω2 , can
ones, especially in millimeter-wave band applications. be used to derive two velocities as v1 = λω1 /(4πTc ) and
FMCW Signal: As previously mentioned, a chirp is a lin- v2 = λω2 /(4πTc ). The velocity resolution, vres , of the radar
early modulated FMCW signal: it is a sinusoidal function is inversely proportional to the duration of a single frame,
formulated as xtx = sin(ωtx t + φtx ), where the frequency Tf = N Tc . By knowing the frame duration Tf , the velocity
ftx = ωtx /(2π) increases linearly over time. After transmis- resolution is vres = λ/(2Tf ) [196].
sion, the reflected chirp signal from an object is collected at Angle Measurement and Resolution: In radar sensing appli-
the RX antenna and can be written as xrx = sin(ωrx t + φrx ). cations, most often the “angle” refers to the horizontal-plane
The intermediate-frequency (IF) signal is produced by mixing AoA at the receiver (or azimuth in a spherical coordinate
RX and TX signals in the mixer component

of the radar as system). It is calculated by observing the phase changes
xif = sin (ωtx −ωrx )t+(φtx −φrx ) . The time delay between occurring on the range-FFT or Doppler-FFT peaks. In order
the RX and the TX signals is τ = 2d/c, where d is the distance to observe these changes, there have to be at least two RX
to the objects and c is the speed of light in air. The start of antennas. The difference between the readings of each antenna
the IF is at τ , which is also when RX chirp is realized and is what produces the phase change in the FFT peaks. The phase
ends when the TX chirp is entirely transmitted. Time delay change is formulated as ∆φ = 2π∆d/λ s.t. ∆d = ` sin(θ),
is the foundation for computing the range and velocity of a where ` is the distance between the antennas. Accordingly,
the angle can be estimated as θ = sin−1 λ∆φ/(2π`) . The


target in an environment. While the given introductory FMCW
concepts are sufficient for the purpose of this paper, further closer θ is to zero, the more accurate the angle estimation
details on FMCW radars can be found in [195]. becomes. In fact, the angle resolution θres = λ/N d cos(θ) is
Range Measurement and Resolution: Range resolution is usually given assuming θ = 0 and d = λ/2 which simplifies
defined as the ability of a radar to identify closely packed it to θres = 2/N . The field of view of the radar depends on
objects. When the distance separating two objects is smaller the maximum AoA that can be measured [196].
than the resolution, the radar becomes unable to distinguish We remark that the distance and angle resolution of a
between them, returning a single range reading. The range mmWave radar device are especially important as they char-
measurement is carried out by computing the phase difference acterise the density and the minimum separation of the points
between TX and RX chirps, yielding the initial phase of an that are detected in the radar maps (see next section). This, in
IF signal, that is formulated as φ0 = 2πfc τ , where fc = turn, has a major impact on the resolution of the clustering al-
c/λ stands for the frequency, c is the speed of light and λ gorithms that are used to separate signals reflected by different
 25

TABLE VI that the carrier frequency is fc , then the phase of the baseband
S UMMARY OF THE HARDWARE PLATFORMS USED IN THE LITERATURE signal is defined as φ(t) = 2πfc 2r(t)c . With this, it is possible
to obtain the micro-Doppler frequency shift caused by the
Hardware platform Related literature
motion of an object. Taking the time derivative of the phase
Google SOLI [197] d
Infineon SiGe [198]
yields dt φ(t) = 2πfc 2c dt
d
r(t). We manipulate this equation
INRAS RadarLog [183] by introducing the Doppler shift induced by the rotation of
Keysight EXG N5172B [199], [200] the object and referring to vector p as the location of an
Qualcomm 802.11ad device [187]
Xilinx Kintex Ultrascale [201]
arbitrary point on it. Thus, the micro-Doppler frequency shift
TI AWR1243 [202] equation is obtained as fD = 2f T
c (v + Ω × p) · n. The first
TI AWR1443 [182], [189], [191], [194] term of the equation is the Doppler shift due to the object’s
TI AWR1443BOOST [191], [203]
TI AWR1642 [204], [205]
translation ftrans = 2fc v · n, where v is the bulk velocity of
TI AWR1642BOOST [186], [206]–[208] the object and n is the radar’s line of sight direction. The
TI AWR1643BOOST [185] second term is the Doppler shift due to the object’s rotation
TI AWR1843BOOST [209]
TI AWR2243 [193]
frot = 2f T
c (Ω × p) · n, where Ω is the angular velocity of the
TI AWR6843 [188], [190] object. In order to get time-varying frequency distribution of
TI IWR1443 [210], [211] micro-Doppler modulation, the short-time Fourier transform
(STFT) [213] is used. STFT is a moving window Fourier
transform where the signal is examined for each window
subjects and objects in the radar maps (see, e.g., density-based interval in order to generate a time-frequency distribution.
spatial clustering of applications with noise (DBSCAN) in the This process can be pictured as a DFT multiplied by the
following sections) and, on the final tracking performance of sliding window’s spectrum, which yields a spectrogram of
any signal processing pipeline. time-varying micro-Doppler modulation [214], [215]. Due to
In Table VI, we summarize the types of passive mmWave the different characteristics in micro-Doppler, it is possible to
radars employed in the literature covered by our survey. detect a moving body and even to identify it, by capturing the
We observe that the availability of commercial evaluation particular modes of motions of the body parts.
boards from Texas Instruments (TI) and of software interfaces Kalman filter (KF) — It is a key tool for the analysis
enabling the retrieval of raw radar data has made TI devices the of time-series containing noise or inaccuracies, providing a
platforms of choice in many of the works. However, others still precise understanding on how the signal changes with time.
prefer powerful but less commercial devices or come up with The KF estimates the state of the monitored process through
custom boards when commercial platforms are not sufficient time, by removing random noise. It is commonly used in
to satisfy the requirements of the application. movement control, navigation and activity recognition, and it
is as well widely employed in radar applications. The discrete
KF (DKF) was initially developed in [216]. It is a two-step
D. Key Processing Techniques recursive algorithm. The first step of the recursive loop is the
Next, we describe some key processing techniques that prediction step, where a projection from the current state of the
are utilized in modern mmWave based radar systems. As model and corresponding uncertainties into the next time-step
detailed below, these are used for various purposes such is made. Second, the correction (or update) step is where
as noise removal, object/people tracking, people detection adjustment of the projection is made by taking the weighted
and identification, vital signal estimation, etc. Note also that average of the projected state with the measured information.
multiple techniques are often used concurrently as part of In linear systems with additive Gaussian noise, DKF works as
the same solution. By processing distance, velocity and angle an optimal least-square error estimator. While for non-linear
information, it is possible to get two or three dimensional systems, the most common KF variants are the EKF and the
data points, such as range-Doppler (RD), range-azimuth (RA) UKF. One of the possible ways of obtaining state estimations
or range-Doppler-azimuth (RDA) maps. This type of data in non-linear models is converting the system into a linear one.
shape, with temporal information between the data frames, At each time step, the EKF uses a first-order partial derivative
can be further processed to provide valuable information about matrix for the evaluation of the next predicted state starting
objects and users in positioning, tracking and identification from the current one. This essentially forces the system to use
applications. linearized versions of the model in the correction step [217].
micro-Doppler — In addition to the main (bulk) movement However, when the model is highly non-linear, the EKF could
of an object, it is possible to have mechanical vibrations experience very slow convergence to the correct solution. In
within the object body as well. These internal vibrations such non-linear models, the KF is used with an unscented
are called micro motions. The micro-Doppler phenomenon is transformation and hence the derivation of UKF. In order to
observed when these micro motions from the object cause carry out predictions, the UKF picks a finite set of points
a frequency modulation on the returned signal [212]. An (called sigma points [218]) around the mean and generates a
example for this would be the individual movements of the new mean by passing this set through the non-linear function
legs and the arms of a person while walking, or rotations of that describes the system. Thus, the new estimate is obtained.
the propellers of a fix-winged aircraft while flying. Assuming While the computational complexity of both filters are same,
that the scalar range from the radar to an object is r(t) and for most cases the UKF practically runs faster as compared to
 26

the EKF, as it does not calculate partial derivatives [219]. algorithms are k-means (based on partitioning), AUTOCLASS
In radar systems, he KF makes it possible to reliably (based on Bayesian statistics), expectation maximization (EM)
estimate the trajectory of the targets, which is achieved by (based on parametric statistics) and also unsupervised neural
filtering the temporal sequences of points in the RD, RA networks and DBSCAN [222] (density based). More on the
or RDA maps, by identifying the center of mass (COM) existing clustering models and algorithms, their categorization
of the moving target(s) and estimating its (their) trajectory and discussion can be found in [223].
(trajectories) over time. KF allows coping with random noise,
obtaining robust trajectories, and to also estimate tracks for DBSCAN — Considering the data gathered by mapping the
the targets in those cases where some temporal RD/RA/RDA radar signal on the environment are tightly packed points in
frames are lost due to occlusions see, e.g., [183]. Given the range, angle and velocity dimensions, one algorithm stands out
sampling time of radar applications and the typical speed of in the field of radar sensing, density-based spatial clustering
movement of people, linear KF models are usually appropriate of applications with noise (DBSCAN) [222]. DBSCAN is a
for human trajectory tracking. Also, most prior works use KF density-based clustering technique where the points belonging
to track the COM of an object or person, treating it as an to a high density region are grouped discarding those that
idealized point-shaped reflector. are recognized to be isolated, in accordance with precise
A recent solution for mmWave indoor radars [184] uses an definitions of the neighborhood of a point and of its local
extended object tracking KF, which makes it possible to jointly density. The algorithm starts at an initial point featuring a
estimate the COM and the extension of the target around it. dense neighborhood and tags it as a core point. The remaining
In [184], such extension is mapped onto an ellipse around the points within the core point’s neighborhood, i.e., within a
COM, whose shape and orientation matches those of the target. preset radius from it, are referred to as reachable. Upon initial-
This KF technology has similar performance as standard KF izing the first core point, DBSCAN evaluates the neighborhood
assuming point-shaped reflectors in terms of tracking accuracy density of each reachable points within its neighborhood, and
for the COM, but additionally it makes it possible to track the ones residing within a dense neighborhood are chosen
the object extension over time. In the case of human sensing, as the new core points. The density connected region of the
the ellipse represents the way the torso is oriented within the cluster is thus extended by connecting dense neighborhoods,
monitored environment. This information, combined with the constructing clusters of generic shapes and only containing
target trajectory, reveals where the target is steering at, which densely connected points. This process is continued in a
may be a valuable information for some applications, e.g., for recurrent fashion until there are no more reachable points
the retail industry. whose neighborhood exceeds the minimum density. Finally, a
cluster is defined as the collection of all points that are either
density-connected or density-reachable. Multiple clusters are
E. Main learning techniques possible and represents density-connected regions of points.
Nowadays, ML and especially DL is successfully being Points that are not contained within a high-density cluster are
applied to many different fields and applications. Although referred to as outliers (these are termed noise points, and are
most of these techniques have been developed for a long time, rejected).
they are recently emerging due to hardware advancements. ML In mmWave based radar applications, DBSCAN has been
methods are used for regression, classification and clustering extensively and successfully used to extract the clusters of
tasks. A more comprehensive analysis and discussion of ML data points in the RD, RA or RDA maps associated with
and DL techniques can be found in [220] and [221], respec- the tracked humans and/or objects (e.g., vehicles) in the
tively. Just like many other fields, these techniques are being monitored environment [224]. This technique was found to
successfully and abundantly exploited within mmWave radar be very robust and efficient due to the following reasons:
sensing systems. i) most importantly, DBSCAN is an unsupervised method,
In some cases, it is required to group sets of objects into its simplest version only needs two parameters to work (a
categories, i.e., to perform clustering. This technique is widely density threshold and a radius for the density neighborhood),
used in such areas as pattern recognition, image analysis while it does not need one to know in advance the number
and machine learning. This is particularly relevant when of clusters (objects/humans) to be tracked. The DBSCAN
there are scattered data points in the observed space, and the parameters are to be set at training time and, for given
information about which point belongs to what category is hardware (mainly, working frequency, distance and angle
non trivial. In our setup, it is used for the analysis of radar resolution) and environment choices (e.g., indoor vs outdoor),
images. After the cluster analysis, if the results are good, the their set values remain rather effective across a large number
clustering method could be exploited to compute labels on of scenarios [183], ii) DBSCAN runs fast, with a time
the dataset, and it could even be used as a part of a more complexity of O(n log n), where n is the number of points
sophisticated system, e.g., for a subsequent identification to be evaluated, iii) the clusters do not have to be spherical,
of the subject or of the human that has generated each DBSCAN works well with clusters of any shape and it is very
data cluster within an image. Often, the clustering task is effective in rejecting random noise, which is quite common
carried out in an unsupervised fashion. Over the years, many in radar maps. Further discussion on how DBSCAN is used
researchers have designed clustering algorithms tailored for in radar systems and applications from the state-of-the-art is
a variety of models. Some of the well known of clustering presented in Section VI-F.
 27

can describe the way a person moves his/her limbs while

Neural networks — The term neural network (NN) performing a certain activity. Vanishing or exploding gradients
comes from biological processes where a collection of are commonly seen during the back-propagation [221] based
neurons create a network. In the modern sense, NNs are training of an RNN. This prevents the NN to effectively
the technology counterpart of the brain. They try to achieve learn, leading to a premature stopping of the training process.
learning by identifying the relationships in a set of data Long-short term memory (LSTM) cells, or alternatively
similarly to how brain does [225]. The most basic NN is the gated recurrent unit (GRU) cells [230], extend the original
perceptron originally devised by Rosenblatt [226]. It only RNN neuron to effectively cope with vanishing or exploding
has a single layer and performs a classification task based gradients [231], by intelligently redefining the structure of a
on taking the input and multiplying it by given weights, memory unit. This solves the gradient vanishing problems at
summing the resulting signals, and passing the result through the cost of a greater complexity.
a non-linear decision function. Essentially, this is the idea Use with mmWave radar signals: activity recognition
behind the whole DL field. Below, we talk briefly about some usually cannot be achieved on data coming from a single
state-of-the-art DL architectures, which have captured the time step, e.g., from a single RD/RA or RDA map. For an
attention of researchers working on radar sensing applications. activity to be determined, analysis of a sequence of such
radar maps should be carried out. Combining this with the
Convolutional NNs (CNNs) — One of the most common micro-Doppler effect observed in humans, it is possible to
NN models is the CNN [227]. CNNs usually consist of estimate the identity of a person based on the specific micro
back to back convolutional and pooling layers with a final motions of their body parts [183], [184], [232].
fully-connected layer. The convolutional layers take the
input and process it via a kernel function (a filter) where Autoencoders (AEs) — Autoencoders [233] encode the input
the feature detection is carried out. These feature maps are and then decode it to generate the output. While an AE is
then fed to the pooling layer where dimension reduction of trained to copy the input onto the output, the main rationale
the domain is performed. This process is continued until behind this is to learn internal representations (features) that
a fully-connected layer, where a flattened feature map is describe the manifold where the high-dimensional input signal
computed and used to obtain the classification output (either resides. That is, the AE features should be highly represen-
via a single non-linearity or a softmax layer). CNN is a tative of the input and can be used to classify it with high
feedforward NN where information can only move in the accuracy. For a proper training of the AE, the encode/decode
forward direction from the input to the output layer, without functionalities are constrained, e.g., by limiting the number of
cycles nor loops. While the convolution operation is naturally neurons in the inner layer or forcing some sparsity for the inner
invariant with respect to rigid translations of input patterns, representation. This forces the AE network to approximate
it does not work with other types of transformations, such as the output by preserving the most relevant features. Denoising
rotations. For this reasons, in the radar sensing field dedicated autoencoders [233] are trained to denoise the input signal
CNN-based approaches have been specifically proposed for and reconstruct, at their ouput, the original (noise-free) signal
radar point clouds, which are discussed shortly below. version. This was found to produce better features in the AE
Use with mmWave radar signals: due to the lack of inner layers. In addition, the denoising capability of such NN
mathematical models to describe RD/RA and RDA maps architectures is valuable for RD/RA/RDA radar maps due to
from mmWave radars and to the presence of strong noise their noisy nature.
components (e.g., from ghost reflections and metallic objects), Use with mmWave radar signals: radar system are
CNN have been extensively used to automatically obtain prone to noisy data and can be significantly affected by
meaningful feature representations from radar readings. unwanted or fake reflections (e.g., ghost reflections). Due
Usually, CNNs are applied to the RD/RA/RDA clusters to this, many radar applications use the AE encode/decode
found by a preceding clustering algorithm, e.g., DBSCAN, functionalities as a middle ground for the reconstruction
assuming that each cluster represents a target object within of the desired observation such as anomaly analysis for
the monitored space. These representations can be then human fall detection [209], person detection for surveillance
utilized to detect objects within an environment [228], to systems [234] and indoor person identification [183].
assess the type of activity a person is carrying out [204], or
to even track their identity [183]. Generative adversarial networks (GANs) — In general,
GANs [235] are divided into two sub-models called the
Recursive NNs (RNNs) — Unlike feedforward NNs, RNNs generator and the discriminator. In the generator network the
[229] utilize their internal memory to retain information from expected outcome is a newly generated sample which should
previous input samples. This allows temporal sequences to reflect the features found in the input data/domain. Conversely
be used as input and thus the learning process can extract the task of the discriminator network is to classify an input
temporal correlation. Hence, RNNs remember the information to detect whether it is a fake (generated) or a real example.
during the learning process, while feedforward NNs cannot. Learning proceeds as a game, where the generator becomes
This is especially relevant for radar data, as it makes it progressively better in generating fakes, and the discriminator
possible to extract temporal features from a sequence of improves at detecting them. The final goal is again to learn
radar maps (i.e., a trajectory). For example, such features meaningful representations (features) of a (usually) high-
 28

dimensional input signal. and [240], novel Pointnet based NNs are presented to track and
Use with mmWave radar signals: Because of the competitive identify people from point clouds obtained by mmWave radars.
nature of the generator and discriminator networks, jumping We remark that mmWave systems can either operate on dense
from one to another during training makes them better at their radar Doppler maps, or on point clouds which can be derived
respective tasks. Most often, algorithms exploit this fact to from these dense maps by only retaining the most significant
generate the required data and use this newly generated input (strongest) reflections. Point clouds are less informative, as
whenever it fits. For instance, GANs have been used in [194] some information is lost when moving from dense to sparse
to generate dense maps from sparse inputs (also referred to representations, but are on the other hand easier to store,
as super resolution imaging) for the purpose of environment transmit and their processing is also lighter. For these reasons,
mapping in a low-visibility environment. In this case, the algorithms that operate on sparse point clouds are particularly
generator network is used to improve the image resolution appealing and are gaining momentum.
and the discriminator to train the generator better.
F. Selected Applications
Residual networks (ResNets) — ResNets [177] use short-
cuts to skip layers. Typically, the skipped layers include Some of the works that we review in this section adopt a
activation and batch normalization [236]. The reason behind custom design for the whole sensing system, from the radar
using shortcuts is to overcome vanishing gradients and/or hardware to the implementation of the software. Others,
gradient degradation problems. Despite the seemingly simple instead, use off-the-shelf radar devices and present new
architectural change, this leads to a major change in terms of algorithms. Most of the applications deal with human activity
learning paradigm, which preserves the correct propagation of recognition, object detection and health monitoring, but
the error gradients across the whole network, allowing one to other use cases are emerging such as vibration detection,
build very deep networks with hundreds of layers and with a environment mapping and even speech recognition.
remarkable representation (feature extraction) effectiveness.
Use with mmWave radar signals: Due to the large number Human Activity Recognition
of layers that can be stacked, ResNets are exploited in For the purpose of tracking and identity recognition
complex scenarios where the input signal contains a high of humans moving in a room, the authors of [183]
number of patterns to be concurrently classified. Examples use micro-Doppler signatures obtained from back-scattered
include human skeletal posture estimation [207], where the mmWave radio signals. An off-the-shelf radar is used to
detection of more than 15 joints and the subsequent tracking gather RDA maps and noise removal is carried out. Hence,
of the person are carried out, or real-time object detection for DBSCAN is applied to the pre-processed RDA maps to detect
autonomous driving [228], where real time obstacle detection the data points (signal reflections) generated by each of the
is performed. human targets in the monitored environment. With DBSCAN,
a cluster of RDA points is obtained for each subject and
PointNet and PointConv — Images are represented through updated as the targets move, across subsequent time steps.
dense regular grids of points, whereas point clouds are irreg- Trajectory detection is carried out by applying a KF to the
ular and also unordered. For these reasons, using the con- clusters and, as a final step, identity recognition is carried out
volution operation with them can be difficult. Pointnet [237] using a CNN with inception layers. The average accuracy is of
is a deep neural network which uses unordered 3D (graph) 90.69% for single targets, 97.96% for two targets, 95.26% for
point clouds as input. The applications of PointNets are object three targets and 98.27% for four targets. Similarly, authors
classification and semantic segmentation. An extension of this in [201] have designed RAPID in order to use off-the-shelf
network is Pointnet++ [238], where the PointNet architecture IEEE 802.11ay devices for person detection and activity recog-
is recursively applied on a nested partitioning of the given nition. Underlying techniques for human activity recognition
point cloud. PointNet++ can identify local features on a greater are similar to the previous work (e.g micro-Doppler signatures,
contextual scale. The key reason of using such architecture KF, CNN). However, RAPID uses CIR estimation and TRN
is to make the extracted features permutation invariant with fields to expose targets movement information from the radio
respect to the input signal. Along the same lines, in [239] signals. As a result, the authors have achieved person detection
the convolution filter is extended into a new operation, called accuracy between 97.8% (for 2 subjects being the highest) and
PointConv, which can be effectively applied to point clouds to 90% (for 7 subjects being the lowest). In addition, activity
build convolutional neural networks. These new network layers recognition rates for walking, running, sitting, and waving
can be used to perform translation-invariant and permutation- hands are 92.9%, 71.6%, 99.8%, and 89.9% respectively.
invariant convolutions (and obtain invariant features) on any Similarly, in [204] micro-Doppler signatures are extracted
point set in the 3D space. These qualities are especially and exploited for human motion detection, where both RD
important for radar point clouds. When tracking people or data cubes as a whole, and RD point clouds are considered.
objects from radar data, being rotation invariant is relevant as The real-time information is received by passing RD data
the traits that we want to capture about the target (movement through Doppler-time extraction. DBSCAN is used to group
of limbs, body shape, etc.) do not depend on their orientation the RD point cloud data from each of the tracked users in the
in space. monitored space. The movements of arms, torso and legs of
Use with mmWave radar signals: In the recent papers [184] a walking person are then identified via a CNN model. Tests
 29

were carried out for walking and leaving, waving hands, sitting RAE had 38% probability of detecting a fall.
to walking transition, walking back and forth, and combining The authors of [203] designed a system to classify static
all behaviors. An average accuracy of 96.32% (walking), hand gestures, namely, palm facing the radar, hand perpen-
99.59% (leaving), 64% (waving hands), 91.18% (sitting to dicular to radar and thumbs-up gesture. In addition to the
walking), 97.84% (walking back and forth) and 95.19% (all) real data, artificial reconstructions of the gestures were used
was observed for each scenario, respectively. to gather synthetic data. Tests were performed both on range
In the same vein, movement pattern detection for of one and RA maps and, 85% and 90% accuracy were respectively
or two patients is the key result in [206]. Together with achieved with them, while with the addition of synthetic data,
DBSCAN, Kalman filtering has been applied to track the the accuracy increased up to 93.1% and 95.4% for range and
trajectory of each patient. Walking, falling, swinging, seizure RA maps, respectively.
and restless movements are the movement patterns which are A framework for human detection and tracking by using
classified by the proposed CNN model. For these movement radar fusion is presented in [191]. Ground truth data is
types, the authors have obtained accuracy values ranging obtained via a camera system. DBSCAN is used for clustering
between 82.77% and 95.74%. and temporal relationships between clusters are exploited to
The authors of [199] have proposed a framework called obtain the probability distribution of the new positions to
“mmSense”. It uses an LSTM-based classification model for perform tracking, similar to Kalman filtering. The concurrent
localization. Initially, environment fingerprinting is carried out use of two radars allows improving the accuracy from 46.9%
with and without human presence. Hence, the presence of to 98.6%.
people and their location within the environment are estimated The “GaitCube” algorithm was proposed in [189]. It utilizes
using an LSTM model. Moreover, an approach combining hu- so called gait data cubes, i.e., 3D joint-feature representations
man outline profile and vital sign measurements gathered from of micro-Doppler and micro-Range signatures for human
60 GHz reflected signal strength series is devised to identify recognition purposes. The idea behind this algorithm is to
the targets. mmSense was tested on five people concurrently exploit the radar’s multi-channel capabilities to improve the
sharing the same physical space, achieving an accuracy of recognition accuracy. Their proposed system achieves 96.1%
97.73% for classification and of 93% for identification tasks, accuracy with a single antenna, 98.3% when using all antennas
respectively. and an average accuracy of 79.1% when tested in an environ-
With the purpose of preventive decision making in au- ment not seen at training time.
tonomous driving applications, the authors of [207] propose Akin to the objectives of the above paper, [188] develops a
“mm-Pose”, a model for estimating the posture of a person in posture estimation algorithm using DBSCAN to cluster and
real-time. To achieve this, RDA data is used to obtain 3D cloud single out real targets. The authors generate their dataset
point representations and red-green-blue (RGB) projections of by installing the radar on the ceiling, and receiving data at
depth-Azimuth and depth-Elevation are used. CNN is used to 10 frames per second. The data processing model is based on
cope with noise and unwanted reflections and also to detect CNNs, and the CNN network is trained on lying, seated and
skeletal joint coordinates. The final model was able to locate upright moving postures. Classification results demonstrate a
17 human skeletal joints with errors of 3.2 cm, 2.7 cm and mean accuracy of 99.1% and an average processing time of
7.5 cm on the depth, elevation and azimuth axes, respectively. 0.13s.
A similar application is presented in [185] for human Another work in [205] performs the classification of 7 fit-
skeletal pose estimation. In this model, range-angle heatmaps ness exercises. CNN and LSTM neural network models are
are initially fed to a CNN followed by a fractionally strided utilized for the classification task, by training them on RD,
convolutional network (FSCN). To exactly locate the target RA, angle-Doppler (AD) and joint-image data. For these data
points, the non-max suppression algorithm was used and the types, a classification accuracy of 92.08%, 98.65%, 97.7%
obtained key points were combined, implementing and testing and 99.27% is attained, respectively. In [182], fitness activities
the proposed solution on single user scenarios. The evaluation were tested both in offline and also in online scenarios.
metrics used in this work are object keypoints similarity Classification was performed on 5 human activities achieving
(OKS) and Mean Average Precision (AP) over different OKS an accuracy of 93.25% and 91.52% for the offline and online
thresholds. The authors obtained an average OKS of 70.5 over operation modes, respectively. The system was also tested
eight different body parts. As a comparison, camera based on multiple locations and the obtained average accuracy is
pose estimators achieve higher performance, i.e., Openpose 88.83%.
(avg. OKS: 93.3) and Leave One Out (avg. OKS: 66.6). Authors of [210], [211] have created a human detection
In [209], a fall detection framework, called “mmFall” is and tracking algorithm by using two radars simultaneously.
presented. 4D cloud points are used, i.e., range, azimuth angle, In both of these works, Kalman filter and DBSCAN were
elevation angle and Doppler. To perform fall detection, the used for tracking and identifying the location of the person,
authors exploit a sequence-2-sequence hybrid variational RNN and the synchronization of the radars were carried out in
autoencoder (HVRAE) model that utilizes an encoder/decoder an offline fashion. The results in [211] show a significant
logic. They use a tailored loss function along with a simplified improvement when a two-radar setup was used with an ac-
version (HVRAE SL) for testing purposes. They also test curacy of 98.6% compared to 46.9% from single-radar setup
the model on vanilla Recurrent Autoencoders (RAE). Overall, in human detection. In [210], radars were placed so that one
HVRAE achieved 98%, HVRAE SL had 60% and vanilla had a top-view and the other had side-view angle. This work
 30

in addition to prior work proposes an alarm system and a background reflections. After this, the RX antenna is steered
posture estimation method. An alarm system is triggered upon according to the peak response observed on the reflected
a positive evaluation in the change of cluster number, number received signal. In other words, the antenna is adaptively
of points in the cluster or the center point of the cluster. The steered to face and track the direction of the pen. Finally,
posture estimation is done for standing, sitting and lying poses the target movement tendency is evaluated by the trend and
by analysing the height of a person at a particular instance and amount of phase shifting. The system can detect the location
the accuracy of estimation is from 92.5% to 93.7%. of the pen at a 94% accuracy, with a tracking error smaller
Towards performing human activity recognition, any than 10 mm across 90% of the trajectory.
combination of range and Doppler (or in some cases of range, In [244], a non-imaging sensor for hidden object detection
Doppler and angle) is used. RDA is typically used with is developed. The authors test their device both in an outdoor
DBSCAN and/or Kalman filtering to identify the clusters scenario with a gun and in an indoor scenario where plastic
within the environment. After extracting micro-Doppler sticks of diameter equal to 2 cm are covered by a fabric. Final
data, a NN architecture (i.e., CNN, RNN, AE etc.) is results of the model are evaluated by applying the Fourier
employed to perform activity/sensing applications. If properly transform to IF chirps to get the range map on horizontal and
designed, DL models are generally the preferred way to vertical scans of the environment compared with a captured
identify movement patterns of RDA clusters, as this is image. In [245], an improved version of the sensor is proposed,
the common theme among most of the surveyed material using a horn antenna integrated with a focusing dielectric lens
above. Deterministic algorithms often fail to provide good operating in the 80–100 GHz frequency range. This sensor can
performance due to the need of a careful parameter tuning be operated with any preferred movement (e.g., up-down) and
(which is very sensitive to the monitored scenario) and to the authors claim that the probability of detection can go up
the lack of mathematical models that accurately represent the to 100%.
signals at the receivers. In [187], an IEEE 802.11ad device is used as a pulsed
wave radar to perform passive handwriting tracking. Slow-
Object Detection and fast-time dimension analysis of the complex CIR, cell-
The authors of [228] propose a method called spatial atten- averaging constant false alarm rate (CA-CFAR) and subsample
tion fusion (SAF) for obstacle detection with mmWave radar peak interpolation (SPI) are the underlying techniques used
and vision sensors. A fully connected one-stage (FCOS) NN is in their algorithm. After applying digital beamforming, the
used for the detection of objects. For the training of this neural authors could extract Doppler maps and by choosing the bins
network, radar data is converted into radar maps (images) with higher Doppler power, could localize the writing tool
and during the feature extraction step, the SAF block within (a pen). Finally, the pen is tracked by picking the lowest
the FCOS network is used for combining radar with vision elevation angle of its lower part at each time-step. With this,
features. The proposed SAF-FCOS model is trained and tested they achieved a tracking accuracy between 3 mm (at a distance
on the nuScenes dataset with a ResNet-50 backbone, achieving of 20 cm) and 40 mm (at a distance of 3 m) and a character
an average precision of 90.0% with an intersection-over-union recognition accuracy ranging from 72% to 82%.
of 0.50 or higher. The authors of [246] perform object classification consid-
The detection of concealed objects implies additional chal- ering three classes: humans, drones and cars. The feature set
lenges, as it becomes necessary to single out hidden objects used in their algorithm consists of radial range, area under the
from rest of the scenery. In [241], the authors employ EM to peak, width of the peak, height and standard deviation of the
fit a Gaussian mixture model of the image acquired: through a peak in the range-FFT domain. Logistic regression and Naive
two-step image segmentation procedure, they first extract the Bayes led to a classification accuracy of 86.9% and 73.9%,
body area from the image and then detect concealed objects. respectively.
The model is evaluated in terms of average probability of error In [192], authors have developed a new deep learning model
and the authors report that multi-level EM has an increased called hybrid dilated convolution (HDC) for concealed object
performance of up to 90.0% when compared to conventional detection. HDC uses two-class semantic segmentation network
EM. for keeping a high resolution in order to detect small objects.
A real-time outdoor hidden object detection model is pro- As a design rule and assignment strategy, expand-contract
posed in [242]. This work also utilizes EM, Bayesian decision dilation (ECD) assignment is applied. In this assignment
making and Gaussian mixture model for image segmentation, stage, the first dilation rate group forms the “rising edge”
with an architecture similar to that of [241]. However, vector (increasing dilation) and the rest forms the “falling edge”
quantization is adopted for the first segmentation level to (decreasing dilation) of dilation rates. As a result, their average
achieve faster computation times. The authors also state that precision with intersection over union of 0.5 is at 0.69% which
EM can be avoided as a whole to reduce computation time outperforms rest of the existing techniques.
(and complexity) significantly, but this causes an error increase As it may be apparent from our discussion, a wide variety of
as well. As a result, [242] achieves a computation time of algorithms have been used for object detection. Initially, signal
1.11 s (with EM) and 0.134 s (without EM) per frame. processing with DFT or FFT is performed to distill signal
Along a similar line, the authors of [243] propose a writing features. Next, such features are either converted into images
object (e.g., a pen) tracking system called “mTrack,” that such as radar maps, or further data processing is applied,
uses dedicated mechanisms to suppress interference from e.g., CIR analysis. ML and DL methods, or decision making
 31

algorithms such as EM, are then applied to obtain the final The authors of [193] use the radars’ multi-channel ca-
results. In general, there is not a single winning methodology. pabilities to improve the estimation of vital signs (heart
Rather, the optimal approach depends on hardware, software, rate). Experiments are performed on 4 different scenarios by
environment and application limitations. changing the location of the radar and the posture of the
Health Monitoring subject. Authors claim that using multi-channel Maximal Ratio
The authors of [247] propose a model for remote heart Combining (MRC) outperforms single channel estimates in
rate (HR) monitoring and analysis. They use the Levenberg- most cases, quantifying the benefits for each scenario.
Marquardt (LM) algorithm for the extraction of heart-rate Although an increasing number of articles is appearing
information. The sum of heartbeat complex, respiration, body on health monitoring via mmWave radars, this field of
movement, background noise, and electronic system noise is application deserves significant additional work. In fact, prior
gathered by expressing the received in-phase and quadrature- art presents results for rather ad-hoc and artificial scenarios,
phase components from LM as the cosine and sine of the where people are still, positioned at known locations, etc. A
received signal. One distinctive advantage of this method is fully automated monitoring system should instead operate
that there is no phase unwrapping as the fitting of the HR in free living conditions, where users are free to move and
signal is directly carried out on the cosine and sine of the no prior location information is available. Further research
received phase modulated signal, which is important for low is thus needed to enable multi-user tracking of vital signs,
SNR scenarios. The method is able to estimate beat-to-beat by also compensating for people movement, which has a
HR and individual heartbeat amplitude, both having a critical detrimental effect on the estimation of breathing and heart rate.
role in the diagnosis of heart diseases.
Other Applications
The authors of [202] demonstrate a remote breathing and
In [186], a system namely “mmVib” for micrometer level
sleep position monitoring system over multiple people at the
vibration detection is presented. The authors propose a
same time. High resolution AoA detection is used to identify
multi-signal consolidation model to understand In-phase and
closely located targets. A support vector machine (SVM) is
Quadrature (IQ) domain and in turn exploit the consistency
used for finding the sleep position, and an optimal filter to
among the two obtained signals to estimate the vibration char-
estimate the breathing rate. The designed system achieved an
acteristics of an object. With this, they can detect vibrations
accuracy of 97% for breathing rate estimation and of 83% for
at micrometer level.
sleep position detection.
The authors of [194] propose an indoor mapping system
In the same vein, the work in [200] proposes vital sign and called “milliMap”, designed for low-visibility environments.
sleep monitoring system. Initially, the location of the person A lidar is used for environment mapping as a ground truth
is detected by using the reflection loss as the classification data collector. A GAN is used to construct the grid map by
parameter, performing a 360° sweep of the environment. recognizing obstacles, free spaces and unknown areas. Finally,
After localizing the human, reflected signal strength samples semantic mapping is applied for the classification of obstacles.
from the reflected signal directed at the human are gathered. A noise-resistant speech sensing framework, “WaveEar,”
For heart rate detection, FFT, customized band-pass filter, is proposed in [249]. Directional beamforming is used to
inverse FFT (IFFT) and peak detection are applied, while for make the system robust to noise. After localizing the throat
breathing rate detection only IFFT and peak detection were and receiving the data, voice reconstruction is achieved by
sufficient. The achieved accuracy was 98.4% and the mean a neural network based on an encoder-decoder (autoencoder)
estimation error in breathing rate and heart rate estimation architecture. As a result, WaveEar achieves a stable 5.5% word
for an incident angle of 70° was smaller than 0.5 bpm and error rate at a distance of about 1.5 m from the user. The
2.5 bpm, respectively. authors also point out that joint optimization speech-to-text
A similar purpose is found in [190], which designs a robot and WaveEar would further enhance the capabilities of their
for human detection and heart rate monitoring. The Hungarian system.
Algorithm and Kalman filtering are used to detect and track the In [208], a mmWave radar device is mounted on a robot
user position. Once the person is located, the robot approaches to estimate its position. This is achieved by exploiting the
him or her and starts the scanning process. The biquad cascade interference produced by other mmWave radars located in
infinite impulse response (IIR) filter is used to extract the the same environment (with known positions), and by only
heartbeat waveforms from the signal, whereas a NN is used estimating the angle of arrival of each other radar interference.
for predicting the heart rate. The proposed system achieved an The proposed system attains position errors for the robot
accuracy between 91.08% and 97.89% across eight different ranging from 14 cm (with three radars) down to 6 cm (ten
poses. radars).
For the purpose of remote glucose level monitoring, the The applications presented in this section vary from
authors of [197], [198] observe reflected multi-channel signal micrometer-level activity recognition to speech recognition.
signatures collected through the SOLI mmWave sensor [248]. We observe that radio sensing enables new and unforeseen
The signal is analyzed by obtaining average power spectral use cases, such as vibration detection [186], indoor naviga-
density (PSD) of each gated signal vector by applying DFT tion [194] and speech reconstruction [249]. However, it is
and FFT. With this, they were able to sense the change in still unclear whether these signals can be reliably detected
dielectric constant due to a varying glucose level in the blood. in an environment with mobility and other sources of noise.
 32

TABLE VII
S UMMARY OF THE ENVIRONMENTS IN WHICH THE EXPERIMENTS HAVE BEEN CARRIED OUT

Evaluation Related literature

Indoors [182], [183], [185]–[194], [197]–[204], [206], [208]–[211], [241], [243]–[245], [247], [249], [250]
Outdoors [185], [205], [207], [228], [242], [244]–[246]

TABLE VIII
S UMMARY OF THE MAIN TECHNIQUES USED IN THE SURVEYED PAPERS

Analytical Tools Related Literature

DBSCAN [182], [183], [188], [191], [204], [210], [211]
Deep learning [182], [182], [183], [185], [188]–[191], [194], [199], [201], [203]–[207], [209], [228], [249]
Fourier transform [186], [187], [197], [198], [200], [203], [208], [244], [246]
Hungarian algorithm [183], [190]
Kalman filter [183], [190], [201], [210], [211]
Levenberg-Marquard method [247]
Machine learning [188], [202], [246]
Non-max suppression [185]
Signal processing [187], [193], [199]–[201], [243], [249]
Statistical modeling [191], [241], [242]

Additional experimental data would be required to check the work is still required. As far as research is concerned, vital
performance of these solutions in general settings and to sign monitoring is still in its infancy as more robust algo-
possibly improve their robustness. rithms are to be developed, capable of working in free living
conditions, i.e., in the presence of user mobility and other
noise sources. In addition, while advanced user tracking and
G. Summary
positioning techniques are available for single radar systems,
In this section, we have summarized the recent advances no substantial work can be found for multi-radar setups, i.e.,
and trends in signal processing for passive mmWave radar radar networks. With multiple radar devices, many additional
systems for indoor spaces. These systems are rapidly gaining problems have to be tackled, including time synchronization,
momentum as radar devices become commercially available, data fusion among radar signals, distributed calibration and
at a low cost. A number of applications are emerging, tar- means to quantify whether and to which extent radar devices
geting diverse scenarios such as people detection, tracking share a common portion of their field of view. For what
and identification, estimation of biosignals such as respiration concerns implementation, much work still has to be performed
and heart rate, detection of gestures/activities/falls, vibrations, architecturally, e.g., where to place the ML based intelligence,
speech or environmental mapping. Table IX summarizes these which messages are to be exchanged between the radars and
application-oriented propositions, while Table VII categorizes the computing units, which protocols are to be exploited to
them based on the environment where the experiments were synchronize multiple devices along time and data dimensions,
carried out. While early works used standard machine learning etc. Lastly, experimental work is key to the development of
algorithms such as expectation maximization and support robust algorithms, as analytical or simulated models often
vector machines, latest developments have been dominated fail to accurately represent all the noise sources. Hence, the
by neural networks. This is clearly evident from Table VIII, collection of experimental data and its publication along with
which presents a summary of the analytical tools discussed the code of the developed solutions are vital to make progress.
in the survey. These are being implemented in their many
flavors, and are allowing researchers to obtain good results
in scenarios where no analytical models are available. As far
as human data monitoring is concerned (e.g., people tracking,
activity monitoring, etc.), the key processing algorithms are
DBSCAN clustering for the separation of user data in the
radar RD/RA/RDA maps and Kalman filtering to reliably track
their trajectories. Neural network architectures are evolving
from standard CNNs to more advanced convolutions (Point-
Conv and PointNets) that were specifically designed for radar
point clouds. Some solutions then use RNNs to capture and
exploit the temporal correlation of radar signals. Advanced
architectures, such as GAN based, are also being exploited to
extract features from radar images.
Although many applications and uses of this technology
have emerged lately, a lot of research and implementation
 33

TABLE IX: S UMMARY OF THE MM WAVE RADAR SENSING WORKS IN THE LITERATURE

Proposition Tools Used Band (GHz) Performance

Human Activity Recognition Algorithms
Micro-Doppler, DBSCAN,
Multi-person tracking and Continuous identification of multiple per-
Kalman filter, Hungarian 77
identification [183] sons with up to 98% accuracy.
algorithm, CNN
Person detection accuracy of 97.8% to
Indoor human detection and CIR, micro-Doppler, Kalman fil- 90%. Walking, running, sitting and waving
60
sensing [201] ter, CNN hands accuracy of 92.9%, 71.6%, 99.8%
and 89.9% respectively.
Multi person classification and identifica-
Multi-person detection and LSTM-based model, RSS series
60 tion accuracy of 97.73% and 93% respec-
identification [199] analysis
tively.
Classification accuracy of 96.1% and
Gait-based human recognition
CNN 77 98.3% with single gait cycle, when using
[189]
single and all receive antenna respectively.
DBSCAN, probability distribu-
Human detection and tracking Human detection sensitivity and precision
tion matching, Kalman filter-like 77-81
[191] of 90% and 98.6% respectively.
algorithm
Offline and real-time activity recognition
Real-time human activity
DBSCAN, CNN, RNN 77 accuracy of 93.25% and 91.52% respec-
recognition [182]
tively, over five different human activities.
Hand gesture classification accuracy of
Hand gesture classification Deep learning, Signal processing
77-81 93% and 95% on range and range-angle
[203] (FFT)
data respectively.
Human motion behavior detec- Accuracy of over 90% in detecting various
Micro-Doppler, DBSCAN, CNN 77
tion [204] human motion behaviours.
Classification accuracy of 92.08%,
Activity recognition and fitness 98.65%, 97.7%, and 99.27% for RD, RA,
Deep learning, CNN 77-81
tracker [205] Angle-Doppler (AD), and joint-image
evaluation respectively.
Over 84.31% prediction accuracy for dif-
Real-time patient behaviour de- ferent behaviors for a single patient.
Micro-Doppler, STFT, CNN 77
tection [206] Around 80% prediction accuracy for dif-
ferent behaviors for two patients.
Detection of 17 human skeletal joints with
Human skeletal pose estima- 3.2 cm, 2.7 cm and 7.5 cm localization
CNN 77
tion [207] error on depth, elevation, and azimuth axes
respectively.
Average object keypoints similarity of 70.5
Human pose estimation [185] CNN, Fractionally strided CNN 77
over 8 different parts.
Proposed scheme achieves 98% fall de-
Fall detection system [209] LSTM, RNN 77 tection rate and outperforms the baseline
techniques.
Real-time posture estimation DBSCAN, CNN, LSTM, Deci- Posture estimation with an accuracy of
77
system [188] sion trees 99.1% at a processing time of 0.13s
Human detection sensitivity of over 90%.
Human detection and track- Two-radar setup improves precision from
DBSCAN, Kalman filters 76-81
ing [210], [211] 46.9% to 98.6%. Posture estimation preci-
sion from 92.5% to 93.7%
Object Detection Algorithms
STFT, CIR, Cell averaging- Tracking accuracy of 3 mm to 40 mm and
Handwriting tracking [187] constant false alarm rate 60 character recognition accuracy of 72% to
(CA-CFAR) 82%.
Obstacle detection for Average precision of 90% with intersection
Deep learning, CNN 77
autonomous driving [228] of unions greater than 0.5.
 34

TABLE IX: S UMMARY OF THE MM WAVE RADAR SENSING WORKS IN THE LITERATURE ( CONTINUED )

Proposition Tools Used Band (GHz) Performance

Gaussian smoothing filter, Usage of multi-level EM increased perfor-
Concealed object
expectation-maximization, 37.47 mance up to 90% compared to conven-
detection [241]
Bayesian tional EM.
Expectation-Maximization,
Real-time concealed object de- Computation time of 1.11 s and 0.134 s
Bayesian decision making, 94
tection [242] with reduced processing.
Gaussian mixture model
Writing object tracking The system tracks/locates a pen with sub-
RSS, phase change analysis 60
(mTrack) [243] centimeter accuracy in 90% of the cases.
Concealed object detection
FFT 80-100 Object detection accuracy up to 100%.
[244], [245]
86.9% and 73.9% classification accuracy
FFT, Logistic regression, Naive
Object classification [246] 77-81 using Logistic Regression and Naive Bayes
Bayes
respectively.
The proposed expand-contract dilation
(ECD) scheme has an average precision
Hidden object detection [192] Semantic segmentation, CNN 60
(AP@0.5) of 0.69, and outperforms all the
existing techniques.
Health Monitoring Algorithms
Remote detection of blood glucose levels
Blood glucose level monitoring
DFT, FFT 57-64 by sensing the change in dielectric constant
[197]
and loss tangent.
Energy-density comparison, Demonstrates accurate identification of
Glucose level detection [198] 57-64
DTFT blood glucose levels.
Human finding accuracy of 98.4% and the
Vital sign and sleep monitoring mean estimation error in breathing rate
RSS, IFFT 60
[200] and heart rate is less then 0.43 Bpm and
2.15 Bpm.
Accuracy of 97% and 83% for breathing
Breathing and sleep position
FFT, DOA, optimum filter, SVM 77-81 rate estimation and sleep position detection
monitoring [202]
respectively.
Proposed signal processing chain signifi-
Arctangent demodulation (AD),
Vital sign monitoring [193] 77-81 cantly improves the heart rate estimation
Maximal ratio combining (MRC)
accuracy in all cases.
Neural networks, Hungarian al- Accuracy of 91.08–97.89% over 8 different
Heart rate sensing [190] 60-64
gorithm, Kalman filter human poses.
Capability of estimating beat-to-beat heart
Heart rate analysis [247] Non-linear Levenberg-Marquardt 94
rate and individual heartbeat amplitude.
Other Algorithms
Map reconstruction error within 0.2 m.
Indoor mapping [194] GAN 77
Obstacle classification accuracy of 90%.
Median amplitude error of 3.4 µm for the
Vibration detection [186] FFT, AoA 77
100 µm amplitude vibration.
Robot position estimation Position estimation of the robot with an
AoA, range and doppler FFT 77
[208] error below 20 cm.
Speech sensing [249] SSNR, Neural network 77 5.5% word error rate around 1.5 m distance
 35

VII. D ISCUSSION AND OPEN RESEARCH DIRECTIONS scenarios, safety, and mission-critical applications), as well
as multiple network optimizations (such as optimal client-AP
Our comprehensive review of the state of the art in mmWave associations, predictive re-association before link breakage due
localization and sensing shows that a sizeable set of contribu- to movement or obstacles, or location-aided beam training and
tions have already covered significant work in this research tracking).
area. Such works show that current mmWave equipment, Regarding passive radar sensing, a number of major ad-
even COTS devices, already offer sufficient opportunities to vancements are envisaged. First, most commercial low-cost
incorporate localization as part of communication processes. radars incorporate linear antenna arrays, which have limited
Moreover, commercial implementations of mmWave radars are detection and tracking capabilities. Bi-dimensional antenna
currently very compact, and cater for precise device-free local- arrays would make it possible to detect higher resolution radar
ization and sensing. However, additional efforts are required images in the 3D space, enabling new uses of this technology
to democratize these tasks and make them natively available (e.g., human gait analysis). Even though commercial mmWave
to vertical applications that rely on mmWave connectivity. radars with enhanced capabilities and 2D antenna arrays are
At the current stage of hardware development, fully-custom becoming available, e.g., TI AWR/IWR radars [251] and
signal processing algorithms only apply to software-defined TI cascaded imaging radar MMWCAS [252] with relatively
radio platforms, where fully-digital transceiver architectures large antenna array size, very little work is available to date
can be available upfront. Conversely, commercial-grade hard- exploiting massive MIMO radars. These would allow high
ware does not give full access to internal signal samples and resolution sensing, which makes it possible to detect finer
measurements, requiring more complex processing and often movements and shapes. Also, most of the available research
yielding limited performance. For example, while theoretical only involves a single radar sensor, whereas radars could be as
analysis predicts millimeter-level device localization accuracy well co-deployed, allowing for large-scale monitoring applica-
and fully digital architectures achieve centimeter-level accu- tions. This will give rise to new opportunities and technical
racy, algorithms for commercial-grade mmWave devices typ- challenges to face, such as new techniques to perform sensor
ically achieve decimeter-level 3rd-quartile localization errors. fusion from multiple radar views, self-calibration algorithms
In this perspective, we conclude that promising research for the distributed radar sensors, transmission and compression
directions in the above field would greatly benefit from new- of radar features from multiple sensing units. Architecturally,
generation standard-compliant mmWave transceiver hardware no clear approach was found on where the supporting com-
that also exposes channel state information to external al- puting facilities are to be located, which messages should
gorithms. While some efforts in this direction have been be sent to them and what is the preferred interaction model
announced, there is still no such platform available on the between the field sensors and the computing units. All of
market. The same observation holds for hybrid beamforming this is of key importance especially for large deployments
architectures. While preliminary works exist that exploit hy- involving multiple sensors. Additional opportunities concern
brid beamforming to improve beam pattern directivity and the combination of mmWave radar systems with camera-based
adaptivity, or to make the 802.11ay SLS operations faster, ones (including thermal cameras), to perform data/feature
these architectures could also help localize mmWave devices extraction and fusion across different sensing domains. Finally,
faster, e.g., by enabling faster angular spectrum scanning. a promising research avenue is to modify commercial off-
Moreover, the field still needs scalable algorithms that flexibly the-shelf communication technology, such as the forthcoming
manage the presence of multiple APs or of several clients IEEE 802.11ay, so that it can double as a passive mmWave
in the same area. These algorithms should work, if possible, radar. This would enable joint communications and passive
with zero initial knowledge of the floor plan and surrounding sensing, potentially without having to deploy a dedicated
area, and ideally estimate the whole environment, including the mmWave radar network. The recent creation of the TGbf task
location of the APs and of all reflective surfaces automatically group (working on research and standardization of WLAN
in a SLAM fashion, in order to relieve the need of input sensing towards the IEEE 802.11bf amendment) testifies the
from the user. Significant research opportunities also exist for interest of the community on these emerging topics.
integrating ML algorithms into location systems. Here, the As a general observation, the research on machine learning
main challenges relate to: relieving the need for extensive methods applied to device-based localization remains limited
training datasets, whose collection requires important efforts; compared to device-free radar-based sensing. For device-based
creating models that transfer well across different environ- localization, machine learning methods find their typical appli-
ments, especially indoors; speeding up the convergence of the cation in fingerprinting approaches. Yet, these schemes require
trained models, e.g., through federated learning, particularly a typically lengthy preliminary measurement effort, which is
when involving heterogenous clients. often deemed excessive or impractical. Conversely, modern
All of the above would be important enablers of a fully mmWave radar systems are both compact and affordable, and
integrated device-based sensing and localization system, for expose a number of features that can be more easily passed
which significant research is still needed. The benefits of on to complex learning and clustering algorithms to map
such a system would be enormous, as the seamless integra- environments, track movement, or estimate the occurrence of
tion of device-based localization and communications would some events of interest. The applicability of machine learning
enable advanced location-based services in multiple domains algorithms to to either field could change if more features
(including but not limited to healthcare, massive IoT, industrial become available, e.g, from multiple digital transceiver ar-
 36

chitectures integrated in the same client. For example, this DAC digital-to-analog converter
would make it possible to use machine learning to increase the DBSCAN density-based spatial clustering of applications with noise
DFT discrete Fourier transform
speed of intermediate localization algorithm steps (e.g., angle DKF discrete KF
computations, ranging and simultaneous distance estimation DL deep learning
among multiple mmWave devices, or joint angle/distance DSP digital signal processor
DTI data transmission interval
estimates based on radio features).
ECD expand-contract dilation
VIII. C ONCLUSIONS EKF extended Kalman filter
EM expectation maximization
Millimeter-wave (mmWave) communication devices will EN-DC enhanced UTRA-dual connectivity
soon become a fundamental component of 5G-and-beyond ESPRIT estimation of signal parameters via rotational invariance tech-
niques
communication networks. This survey put the lens on re-
cent research advances in localization and sensing algorithms FCOS fully connected one-stage
for indoor mmWave communication and radar devices. Af- FFT fast Fourier transform
FMCW frequency-modulated continuous wave
ter introducing the most important properties of mmWave FPGA field-programmable gate array
signal propagation and communication chain architectures FSCN fractionally strided convolutional network
that enable mmWave channel measurements, we presented FTM fine time measurement
a thorough account of localization algorithms for mmWave
GAN generative adversarial network
devices. These are based on a broad range of techniques, GRU gated recurrent unit
that include both traditional methods based, e.g., on timing GSCM geometry-based stochastic channel model
and received signal strength indicator (RSSI) information, and
HDC hybrid dilated convolution
more specific methods that exploit the properties of mmWave HR heart rate
devices and signal propagation, e.g., by processing channel HVRAE hybrid variational RNN autoencoder
state information (CSI).
IF intermediate-frequency
Then, we turned our attention to consumer-grade mmWave
IFFT inverse FFT
radar devices, which are becoming extremely cost-effective IIR infinite impulse response
sensing platforms. After introducing the basic structure of such IoT Internet of things
radar architectures, we discussed different approaches that ITU-R International telecommunication union – radiocommunication
Sector
tackle applications such as human activity recognition, object
detection and health monitoring. We unveiled that several KF Kalman filter
research directions remain open in both fields, including better
LM Levenberg-Marquardt
algorithms for localization and sensing with consumer-grade LMS least mean squares
devices, data fusion methods for dense deployments, as well LoS line-of-sight
as an educated application of machine learning methods to LSTM long-short term memory
both device-based localization and device-free sensing.
MAC medium access control
MCU micro-controller unit
L IST OF ABBREVIATIONS MF matched filter
MIMO multiple-input multiple-output
MISO multiple-input single-output
3GPP Third-generation partnership project ML machine learning
4G fourth-generation mmWave millimeter-wave
5G fifth-generation MPC multipath component
MUSIC multiple signal classification
A-BFT association beamforming training
ABT asymmetric beamforming training NLoS non-line-of-sight
AD angle-Doppler NN neural network
ADC analog-to-digital converter
ADoA angle difference-of-arrival OFDM orthogonal frequency-division multiplexing
AE autoencoder OKS object keypoints similarity
AoA angle of arrival
AoD angle of departure PA power amplifier
AP access point PDP power-delay profile
API application program interface PHY physical layer
PRS positioning reference signal
BI beacon interval PSD power spectral density
BRP beam refinement protocol PW pulsed wave
BS base station
RA range-azimuth
CA-CFAR cell-averaging constant false alarm rate RD range-Doppler
CBAP contention based access period RDA range-Doppler-azimuth
CIR channel impulse response ResNet residual network
CNN convolutional NN RF radio frequency
COM center of mass RGB red-green-blue
COTS commercial off-the-shelf RMS root mean square
CSI channel state information RMSE root mean square error
 37

RNN recursive NN [9] L. Subrt, P. Pechac, and S. Zvanovec, “New approach to modeling of
RSS received signal strength diffuse reflection and scattering for millimeter-wave systems in indoor
RSSI received signal strength indicator scenarios,” PIERS Online, vol. 6, no. 8, pp. 719–722, 2010.
RTT round-trip time [10] G. R. MacCartney, J. Zhang, S. Nie, and T. S. Rappaport, “Path
RX receive/receiver loss models for 5G millimeter wave propagation channels in urban
RZF regularized zero-forcing microcells,” in Proc. IEEE GLOBECOM, Dec. 2013, pp. 3948–3953.
[11] M. Giordani, M. Mezzavilla, and M. Zorzi, “Initial access in 5G
SAF spatial attention fusion mmWave cellular networks,” IEEE Commun. Mag., vol. 54, no. 11,
SAGE space-alternating generalized expectation maximization pp. 40–47, Nov. 2016.
SDR software-defined radio [12] Y. Li, J. G. Andrews, F. Baccelli, T. D. Novlan, and J. C. Zhang,
SLAM simultaneous localization and mapping “Performance analysis of millimeter-wave cellular networks with two-
SLS sector-level sweep stage beamforming initial access protocols,” in Proc. Asilomar conf.
SNR signal-to-noise ratio on signals syst. and comput., Nov. 2016, pp. 1171–1175.
SoC system-on-chip [13] F. Devoti, I. Filippini, and A. Capone, “Facing the millimeter-wave
SP service period cell discovery challenge in 5G networks with context-awareness,” IEEE
SPI subsample peak interpolation Access, vol. 4, pp. 8019–8034, 2016.
SRS sounding reference signal [14] T. Nitsche, C. Cordeiro, A. B. Flores, E. W. Knightly, E. Perahia, and
SSW sector sweep J. C. Widmer, “IEEE 802.11ad: directional 60 GHz communication for
STA station multi-gigabit-per-second Wi-Fi,” IEEE Commun. Mag., vol. 52, no. 12,
STFT short-time Fourier transform pp. 132–141, 2014, invited paper.
SVM support vector machine [15] IEEE, “IEEE Standard for Information technology–
Telecommunications and information exchange between systems–
TDoA time difference-of-arrival Local and metropolitan area networks–Specific requirements-Part 11:
Wireless LAN Medium Access Control (MAC) and Physical Layer
ToA time of arrival
(PHY) Specifications Amendment 3: Enhancements for Very High
ToF time of flight
Throughput in the 60 GHz Band,” Std 802.11ad-2012 (Amendment to
TX transmit/transmitter
IEEE Std 802.11-2012, as amended by IEEE Std 802.11ae-2012 and
TXSS transmit sweep IEEE Std 802.11aa-2012), pp. 1–628, 2012.
[16] P. Zhou, K. Cheng, X. Han, X. Fang, Y. Fang, R. He, Y. Long, and
UAV unmanned aerial vehicle Y. Liu, “IEEE 802.11ay-based mmWave WLANs: Design challenges
UE user equipment and solutions,” IEEE Commun. Surveys Tuts., vol. 20, no. 3, pp. 1654–
UKF unscented Kalman filter 1681, 2018.
URLLC ultra-reliable low-latency communications [17] Y. Ghasempour, C. R. C. M. da Silva, C. Cordeiro, and E. W. Knightly,
USRP universal software radio peripheral “IEEE 802.11ay: Next-generation 60 GHz communication for 100 Gb/s
Wi-Fi,” IEEE Commun. Mag., vol. 55, no. 12, pp. 186–192, 2017.
V2X vehicle-to-everything [18] C. De Lima, D. Belot, R. Berkvens, A. Bourdoux, D. Dardari,
VCO voltage-controlled oscillator M. Guillaud, M. Isomursu, E.-S. Lohan, Y. Miao, A. N. Barreto,
M. R. K. Aziz, J. Saloranta, T. Sanguanpuak, H. Sarieddeen, G. Seco-
WLAN wireless LAN Granados, J. Suutala, T. Svensson, M. Valkama, B. Van Liempd, and
H. Wymeersch, “Convergent communication, sensing and localization
ZF zero-forcing in 6G systems: An overview of technologies, opportunities and chal-
lenges,” IEEE Access, vol. 9, pp. 26 902–26 925, 2021.
[19] Z. Xiao and Y. Zeng, “An overview on integrated localization and
R EFERENCES communication towards 6G,” pp. 1–46, Dec. 2021.
[20] F. Lemic, J. Martin, C. Yarp, D. Chan, V. Handziski, R. Brodersen,
[1] M. Xiao, S. Mumtaz, Y. Huang, L. Dai, Y. Li, M. Matthaiou, G. K. G. Fettweis, A. Wolisz, and J. Wawrzynek, “Localization as a feature
Karagiannidis, E. Björnson, K. Yang, I. Chih-Lin, and A. Ghosh, of mmWave communication,” in Proc. IWCMC, 2016, pp. 1033–1038.
“Millimeter wave communications for future mobile networks (guest [21] A. Vashist, M. P. Li, A. Ganguly, S. M. Pd, C. Hochgraf, R. Ptucha,
editorial), Part I,” IEEE J. Sel. Areas Commun., vol. 35, no. 7, pp. A. Kwasinski, and M. E. Kuhl, “KF-Loc: A Kalman filter and ma-
1425–1431, 2017. chine learning integrated localization system using consumer-grade
[2] X. Wang, L. Kong, F. Kong, F. Qiu, M. Xia, S. Arnon, and G. Chen, millimeter-wave hardware,” IEEE Consum. Electron. Mag., 2022, in
“Millimeter wave communication: A comprehensive survey,” IEEE press.
Commun. Surveys Tuts., vol. 20, no. 3, pp. 1616–1653, 2018.
[22] C. de Lima, D. Belot, R. Berkvens, A. Bourdoux, D. Dardari, M. Guil-
[3] A. N. Uwaechia and N. M. Mahyuddin, “A comprehensive survey
laud, M. Isomursu, E.-S. Lohan, Y. Miao, A. Barreto, M. Aziz,
on millimeter wave communications for fifth-generation wireless net-
J. Saloranta, T. Sanguanpuak, H. Sarieddeen, G. Seco-Granados, J. Su-
works: Feasibility and challenges,” IEEE Access, vol. 8, pp. 62 367–
utala, T. Svensson, M. Valkama, H. Wymeersch, and B. van Liempd,
62 414, 2020.
“6G white paper on localization and sensing,” 2020, arXiv preprint
[4] Z. Abu-Shaban, X. Zhou, T. Abhayapala, G. Seco-Granados, and
2006.01779.
H. Wymeersch, “Error bounds for uplink and downlink 3D localization
in 5G millimeter wave systems,” IEEE Trans. Wireless Commun., [23] F. Zafari, A. Gkelias, and K. K. Leung, “A survey of indoor localization
vol. 17, no. 8, pp. 4939–4954, 2018. systems and technologies,” IEEE Commun. Surveys Tuts., vol. 21, no. 3,
[5] F. Ghaseminajm, Z. Abu-Shaban, S. S. Ikki, H. Wymeersch, and C. R. pp. 2568–2599, 2019.
Benson, “Localization error bounds for 5G mmWave systems under I/Q [24] T. Kim Geok, K. Zar Aung, M. Sandar Aung, M. Thu Soe, A. Abdaziz,
imbalance,” IEEE Trans. Veh. Technol., vol. 69, no. 7, pp. 7971–7975, C. Pao Liew, F. Hossain, C. P. Tso, and W. H. Yong, “Review of indoor
2020. positioning: Radio wave technology,” MDPI Applied Sciences, vol. 11,
[6] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N. no. 1, pp. 1–44, 2021.
Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave [25] K. Ngamakeur, S. Yongchareon, J. Yu, and S. U. Rehman, “A survey
mobile communications for 5G cellular: It will work!” IEEE Access, on device-free indoor localization and tracking in the multi-resident
vol. 1, pp. 335–349, 2013. environment,” ACM Comput. Surv., vol. 53, no. 4, pp. 1–29, Sep. 2020.
[7] T. S. Rappaport, F. Gutierrez, E. Ben-Dor, J. N. Murdock, Y. Qiao, and [26] N. Singh, S. Choe, and R. Punmiya, “Machine learning based indoor
J. I. Tamir, “Broadband millimeter-wave propagation measurements localization using Wi-Fi RSSI fingerprints: An overview,” IEEE Access,
and models using adaptive-beam antennas for outdoor urban cellular vol. 9, pp. 127 150–127 174, 2021.
communications,” IEEE Trans. Antennas Propag., vol. 61, no. 4, pp. [27] R. C. Shit, S. Sharma, D. Puthal, P. James, B. Pradhan, A. v. Moorsel,
1850–1859, Apr. 2013. A. Y. Zomaya, and R. Ranjan, “Ubiquitous localization (UbiLoc): A
[8] N. Peinecke, H. Doehler, and B. Korn, “Phong-like lighting for MMW survey and taxonomy on device free localization for smart world,” IEEE
radar simulation,” in Proc. SPIE, 8 2008, pp. 1–10. Commun. Surveys Tuts., vol. 21, no. 4, pp. 3532–3564, 2019.
 38

[28] F. Wen, H. Wymeersch, B. Peng, W. P. Tay, H. C. So, and D. Yang, “A [50] S. Hur, S. Baek, B. Kim, Y. Chang, A. F. Molisch, T. S. Rappaport,
survey on 5G massive MIMO localization,” Digital Signal Processing, K. Haneda, and J. Park, “Proposal on millimeter-wave channel mod-
vol. 94, pp. 21–28, 2019. eling for 5G cellular system,” IEEE J. Sel. Topics Signal Process.,
[29] C. Laoudias, A. Moreira, S. Kim, S. Lee, L. Wirola, and C. Fischione, vol. 10, pp. 454–469, Apr. 2016.
“A survey of enabling technologies for network localization, tracking, [51] S. Jaeckel, L. Jiang, V. Jungnickel, L. Thiele, C. Jandura, G. Som-
and navigation,” IEEE Commun. Surveys Tuts., vol. 20, no. 4, pp. 3607– merkorn, and C. Schneider, “Correlation properties of large and small-
3644, 2018. scale parameters from multicell channel measurements,” in Proc. Eu-
[30] J. Yang, J. Xu, X. Li, S. Jin, and B. Gao, “Integrated communication CAP, 2009, pp. 3406–3410.
and localization in millimeter-wave systems,” Frontiers of Information [52] A. F. Molisch, A. Karttunen, S. Hur, J. Park, and J. Zhang, “Spatially
Tech. & Electronic Eng., vol. 22, no. 4, pp. 457–470, 2021. consistent pathloss modeling for millimeter-wave channels in urban
[31] M. Giordani, M. Polese, M. Mezzavilla, S. Rangan, and M. Zorzi, environments,” in Proc. EuCAP, Apr. 2016, pp. 1–5.
“Toward 6G networks: Use cases and technologies,” IEEE Commun. [53] L. Raschkowski, P. Kyösti, K. Kusume, and T. Jämsä, “METIS
Mag., vol. 58, no. 3, pp. 55–61, 2020. deliverable D1.4: Channel models,” pp. 1–220, 2015, accessed:
[32] T. Wild, V. Braun, and H. Viswanathan, “Joint design of communication Sept. 2021. [Online]. Available: https://metis2020.com/wp-content/
and sensing for beyond 5G and 6G systems,” IEEE Access, vol. 9, pp. uploads/deliverables/METIS D1.4 v1.0.pdf
30 845–30 857, 2021. [54] G. R. MacCartney, M. K. Samimi, and T. S. Rappaport, “Exploiting
[33] A. Bourdoux, A. N. Barreto, B. van Liempd, C. de Lima, D. Dardari, directionality for millimeter-wave wireless system improvement,” in
D. Belot, E.-S. Lohan, G. Seco-Granados, H. Sarieddeen, H. Wymeer- Proc. IEEE ICCW, 2015, pp. 2416–2422.
sch et al., “6g white paper on localization and sensing,” arXiv preprint [55] T. A. Thomas, H. C. Nguyen, G. R. MacCartney, and T. S. Rappaport,
arXiv:2006.01779, 2020. “3D mmWave channel model proposal,” in Proc. IEEE VTC-Fall, Sep.
[34] J. Sanusi, O. Oshiga, S. Thomas, S. Idris, S. Adeshina, and A. M. 2014, pp. 1–6.
Abba, “A review on 6G wireless communication systems: Localization [56] J. Ko, Y.-J. Cho, S. Hur, T. Kim, J. Park, A. F. Molisch, K. Haneda,
and sensing,” in Proc. ICMEAS, 2021, pp. 1–5. M. Peter, D.-J. Park, and D.-H. Cho, “Millimeter-wave channel mea-
[35] H. Tataria, K. Haneda, A. F. Molisch, M. Shafi, and F. Tufvesson, surements and analysis for statistical spatial channel model in in-
“Standardization of propagation models for terrestrial cellular systems: building and urban environments at 28 GHz,” IEEE Trans. Wireless
A historical perspective,” International J. of Wireless Inf. Netw., vol. 28, Commun., vol. 16, pp. 5853–5868, Sep. 2017.
pp. 20–44, Mar. 2021. [57] M.-D. Kim, J. Liang, J. Lee, J. Park, and B. Park, “Directional
[36] M. Shafi, J. Zhang, H. Tataria, A. F. Molisch, S. Sun, T. S. Rappaport, multipath propagation characteristics based on 28 GHz outdoor channel
F. Tufvesson, S. Wu, and K. Kitao, “Microwave vs. millimeter-wave measurements,” in Proc. EuCAP, Apr. 2016, pp. 1–5.
propagation channels: Key differences and impact on 5G cellular [58] M. Shafi, A. F. Molisch, P. J. Smith, T. Haustein, P. Zhu, P. D. Silva,
systems,” IEEE Commun. Mag., vol. 56, pp. 14–20, Dec. 2018. F. Tufvesson, A. Benjebbour, and G. Wunder, “5G: A tutorial overview
[37] H. J. Liebe, “An updated model for millimeter wave propagation in of standards, trials, challenges, deployment, and practice,” IEEE J. Sel.
moist air,” Radio Science, vol. 20, pp. 1069–1089, Sep. 1985. Areas Commun., vol. 35, pp. 1201–1221, Jun. 2017.
[38] J. Kunisch and J. Pamp, “An ultra-wideband space-variant multipath [59] 3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,”
indoor radio channel model,” in Proc. IEEE UWBST, Nov. 2003, pp. 3rd Generation Partnership Project (3GPP), Tech. Rep. 38.901, vol. 14,
290–294. 2018.
[39] J. Poutanen, F. Tufvesson, K. Haneda, V.-M. Kolmonen, and [60] M. Series, “Guidelines for evaluation of radio interface technologies
P. Vainikainen, “Multi-link MIMO channel modeling using geometry- for IMT-2020,” Report ITU, p. 2412, 2017.
based approach,” IEEE Trans. Antennas Propag., vol. 60, pp. 587–596, [61] J. Poutanen, K. Haneda, L. Liu, C. Oestges, F. Tufvesson, and
Feb. 2012. P. Vainikainen, “Parameterization of the COST 2100 MIMO channel
[40] A. Maltsev, A. Pudeyev, A. Lomayev, and I. Bolotin, “Channel model- model in indoor scenarios,” in Proc. EuCAP, 2011, pp. 3606–3610.
ing in the next generation mmWave Wi-Fi: IEEE 802.11ay standard,” [62] S. Jaeckel, L. Raschkowski, K. Börner, L. Thiele, F. Burkhardt, and
in Proc. European Wireless, May 2016, pp. 1–8. E. Eberlein, “Quadriga-quasi deterministic radio channel generator,
[41] W. Keusgen, R. J. Weiler, M. Peter, M. Wisotzki, and B. Göktepe, user manual and documentation,” Fraunhofer Heinrich Hertz Institute,
“Propagation measurements and simulations for millimeter-wave mo- Tech. Rep. v2. 4.0, 2020.
bile access in a busy urban environment,” in Proc. IRMMW-THz, Sep. [63] S. Ju, Y. Xing, O. Kanhere, and T. S. Rappaport, “3-D statistical
2014, pp. 1–3. indoor channel model for millimeter-wave and sub-terahertz bands,”
[42] H. C. Nguyen, I. Rodriguez, T. B. Sorensen, L. L. Sanchez, I. Kovacs, in Proc. IEEE GLOBECOM, 2020, pp. 1–7.
and P. Mogensen, “An empirical study of urban macro propagation at [64] P. F. M. Smulders, “Statistical characterization of 60-GHz indoor radio
10, 18 and 28 GHz,” in Proc. IEEE VTC-Spring, May 2016, pp. 1–5. channels,” IEEE Trans. Antennas Propag., vol. 57, pp. 2820–2829, Oct.
[43] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, 2009.
“Wideband millimeter-wave propagation measurements and channel [65] F. Erden, O. Ozdemir, and I. Guvenc, “28 GHz mmWave channel
models for future wireless communication system design,” IEEE Trans. measurements and modeling in a library environment,” in Proc. IEEE
Commun., vol. 63, no. 9, pp. 3029–3056, 2015. RWS, 2020, pp. 52–55.
[44] S. Hur, Y.-J. Cho, T. Kim, J. Park, A. F. Molisch, K. Haneda, [66] Y. Xing, O. Kanhere, S. Ju, and T. S. Rappaport, “Indoor wireless
and M. Peter, “Wideband spatial channel model in an urban cellular channel properties at millimeter wave and sub-terahertz frequencies,”
environments at 28 GHz,” in Proc. EuCAP, Apr. 2015, pp. 1–5. in Proc. IEEE GLOBECOM, Dec. 2019, pp. 1–6.
[45] R. J. Weiler, M. Peter, W. Keusgen, A. Maltsev, I. Karls, A. Pudeyev, [67] J. Huang, C.-X. Wang, H. Chang, J. Sun, and X. Gao, “Multi-frequency
I. Bolotin, I. Siaud, and A.-M. Ulmer-Moll, “Quasi-deterministic multi-scenario millimeter wave MIMO channel measurements and
millimeter-wave channel models in miweba,” EURASIP J. on Wireless modeling for B5G wireless communication systems,” IEEE J. Sel.
Commun. and Networking, vol. 2016, p. 84, Mar. 2016. Areas Commun., vol. 38, no. 9, pp. 2010–2025, 2020.
[46] C. Gustafson, F. Tufvesson, S. Wyne, K. Haneda, and A. F. Molisch, [68] R. J. Weiler, M. Peter, W. Keusgen, K. Sakaguchi, and F. Undi,
“Directional analysis of measured 60 GHz indoor radio Channels using “Environment induced shadowing of urban millimeter-wave access
SAGE,” in Proc. IEEE VTC-Spring, May 2011, pp. 1–5. links,” IEEE Wireless Commun. Lett., vol. 5, pp. 440–443, Aug. 2016.
[47] C. Schneider, M. Narandzic, M. Käske, G. Sommerkorn, and R. S. [69] J. Medbo, H. Asplund, and J.-E. Berg, “60 GHz channel direc-
Thomä, “Large scale parameter for the WINNER II channel model at tional characterization using extreme size virtual antenna array,” in
2.53 GHz in urban macro cell,” in Proc. IEEE VTC, May 2010, pp. Proc. IEEE PIMRC, Aug. 2015, pp. 176–180.
1–5. [70] W. Fu, J. Hu, and S. Zhang, “Frequency-domain measurement of
[48] P. B. Papazian, C. Gentile, K. A. Remley, J. Senic, and N. Golmie, 60 GHz indoor channels: a measurement setup, literature data, and
“A radio channel sounder for mobile millimeter-wave communications: analysis,” IEEE Instrum. Meas. Mag., vol. 16, pp. 34–40, Apr. 2013.
System implementation and measurement assessment,” IEEE Trans. [71] K. Haneda, J. Järveläinen, A. Karttunen, M. Kyrö, and J. Putkonen,
Microw. Theory Tech., vol. 64, pp. 2924–2932, Sep. 2016. “A statistical spatio-temporal radio channel model for large indoor
[49] K. Haneda, N. Omaki, T. Imai, L. Raschkowski, M. Peter, and environments at 60 and 70 GHz,” IEEE Trans. Antennas Propag.,
A. Roivainen, “Frequency-agile pathloss models for urban street vol. 63, pp. 2694–2704, Jun. 2015.
canyons,” IEEE Trans. Antennas Propag., vol. 64, pp. 1941–1951, May [72] G. R. Maccartney, T. S. Rappaport, S. Sun, and S. Deng, “Indoor of-
2016. fice wideband millimeter-wave propagation measurements and channel
 39

models at 28 and 73 GHz for ultra-dense 5G wireless networks,” IEEE [94] A. F. Molisch, V. V. Ratnam, S. Han, Z. Li, S. L. H. Nguyen, L. Li,
Access, vol. 3, pp. 2388–2424, 2015. and K. Haneda, “Hybrid beamforming for massive MIMO: A survey,”
[73] M. H. Tariq, I. Chondroulis, P. Skartsilas, N. Babu, and C. B. Papadias, IEEE Commun. Mag., vol. 55, pp. 134–141, Sep. 2017.
“mmWave massive MIMO channel measurements for fixed wireless [95] O. E. Ayach, R. W. Heath, S. Abu-Surra, S. Rajagopal, and Z. Pi, “Low
and smart city applications,” in Proc. IEEE PIMRC, 2020, pp. 1–6. complexity precoding for large millimeter wave MIMO systems,” in
[74] Y. Xing, T. S. Rappaport, and A. Ghosh, “Millimeter wave and sub- Proc. IEEE ICC, Jun. 2012, pp. 3724–3729.
THz indoor radio propagation channel measurements, models, and [96] T. Kim, J. Park, J.-Y. Seol, S. Jeong, J. Cho, and W. Roh, “Tens of
comparisons in an office environment,” IEEE Commun. Lett., vol. 25, Gbps support with mmWave beamforming systems for next generation
no. 10, pp. 3151–3155, 2021. communications,” in Proc. IEEE GLOBECOM, Dec. 2013, pp. 3685–
[75] C. Cano, G. H. Sim, A. Asadi, and X. Vilajosana, “A channel mea- 3690.
surement campaign for mmWave communication in industrial settings,” [97] J. Song, J. Choi, and D. J. Love, “Codebook design for hybrid
IEEE Trans. Wireless Commun., vol. 20, no. 1, pp. 299–315, 2020. beamforming in millimeter wave systems,” in Proc. IEEE ICC, Jun.
[76] C. Gustafson, K. Haneda, S. Wyne, and F. Tufvesson, “On mm-Wave 2015, pp. 1298–1303.
Multipath Clustering and Channel Modeling,” IEEE Trans. Antennas [98] S. Payami, M. Shariat, M. Ghoraishi, and M. Dianati, “Effective rf
Propag., vol. 62, pp. 1445–1455, Mar. 2014. codebook design and channel estimation for millimeter wave commu-
[77] S. Salous, V. Degli Esposti, F. Fuschini, R. S. Thomae, R. Mueller, nication systems,” in Proc. IEEE ICCW, Jun. 2015, pp. 1226–1231.
D. Dupleich, K. Haneda, J.-M. Molina Garcia-Pardo, J. Pascual Garcia, [99] S. Dutta, C. N. Barati, D. Ramirez, A. Dhananjay, J. F. Buckwalter,
D. P. Gaillot, S. Hur, and M. Nekovee, “Millimeter-wave propagation: and S. Rangan, “A case for digital beamforming at mmWave,” IEEE
Characterization and modeling toward fifth-generation systems,” IEEE Trans. Wireless Commun., vol. 19, pp. 756–770, Feb. 2020.
Antennas Propag. Mag., vol. 58, pp. 115–127, Dec. 2016. [100] B. Yang, Z. Yu, J. Lan, R. Zhang, J. Zhou, and W. Hong, “Digital
[78] S. Hur, H. Yu, J. Park, W. Roh, C. U. Bas, R. Wang, and A. F. Molisch, beamforming-based massive MIMO transceiver for 5G millimeter-wave
“Feasibility of mobility for millimeter-wave systems based on channel communications,” IEEE Trans. Microw. Theory Tech., vol. 66, pp.
measurements,” IEEE Commun. Mag., vol. 56, pp. 56–63, Jul. 2018. 3403–3418, Jul. 2018.
[79] D. Dupleich, R. Müller, S. Skoblikov, C. Schneider, J. Luo, M. Boban, [101] Y. Hu, “A digital multibeam array with wide scanning angle and
G. D. Galdo, and R. Thomä, “Multi-band characterization of path-loss, enhanced beam gain for millimeter-wave massive MIMO applications,”
delay, and angular spread in V2V links,” in Proc. IEEE PIMRC, Sep. IEEE Trans. Antennas Propag., vol. 66, p. 11, 2018.
2018, pp. 85–90. [102] P. Xingdong, H. Wei, Y. Tianyang, and L. Linsheng, “Design and
[80] F. Fuschini, S. Häfner, M. Zoli, R. Müller, E. M. Vitucci, D. Dupleich, implementation of an active multibeam antenna system with 64 RF
M. Barbiroli, J. Luo, E. Schulz, V. Degli-Esposti et al., “Item level channels and 256 antenna elements for massive MIMO application in
characterization of mm-wave indoor propagation,” EURASIP J. on 5G wireless communications,” China Commun., vol. 11, pp. 16–23,
Wireless Commun. and Networking, vol. 2016, no. 1, pp. 1–12, 2016. Nov. 2014.
[103] C. Yu, J. Jing, H. Shao, Z. H. Jiang, P. Yan, X.-W. Zhu, W. Hong, and
[81] H. Tataria and F. Tufvesson, “Impact of spatially consistent channels
A. Zhu, “Full-angle digital predistortion of 5G millimeter-wave massive
on digital beamforming for millimeter-wave systems,” in Proc. EuCAP,
MIMO transmitters,” IEEE Trans. Microw. Theory Tech., vol. 67, pp.
Mar. 2020, pp. 1–5.
2847–2860, Jul. 2019.
[82] M. Steinbauer, A. Molisch, and E. Bonek, “The double-directional
[104] 3GPP, “Study on channel model for frequency bands above 6 GHz,”
radio channel,” IEEE Trans. Antennas Propag., vol. 43, pp. 51–63,
3rd Generation Partnership Project (3GPP), Tech. Rep. 38.900, vol. 15,
Aug. 2001.
2018.
[83] L. Liu, C. Oestges, J. Poutanen, K. Haneda, P. Vainikainen, F. Quitin, [105] ——, “Study on NR vehicle-to-everything,” 3rd Generation Partner-
F. Tufvesson, and P. D. Doncker, “The COST 2100 MIMO channel ship Project (3GPP), Tech. Rep. 37.885, vol. 16, 2019.
model,” IEEE Wireless Commun. Mag., vol. 19, pp. 92–99, Dec. 2012.
[106] ——, “Study on NR positioning enhancements,” 3rd Generation Part-
[84] K. Haneda et al., “5G 3GPP-like channel models for outdoor urban nership Project (3GPP), Tech. Rep., vol. 38.857, March 2021.
microcellular and macrocellular environments,” in Proc. IEEE VTC- [107] RP-213588, “Study on expanded and improved NR positioning,” Dec
Spring, May 2016, pp. 1–7. 2021, accessed: Mar 2022. [Online]. Available: https://www.3gpp.org/
[85] S. Ju, Y. Xing, O. Kanhere, and T. S. Rappaport, “Millimeter wave ftp/tsg ran/TSG RAN/TSGR 94e/Docs/RP-213588
and sub-terahertz spatial statistical channel model for an indoor office [108] X. Li, E. Leitinger, M. Oskarsson, K. Astrom, and F. Tufvesson,
building,” IEEE J. Sel. Areas Commun., vol. 39, no. 6, pp. 1561–1575, “Massive MIMO-based localization and mapping exploiting phase
2021. information of multipath components,” IEEE Trans. Wireless Commun.,
[86] F. Khan and Z. Pi, “mmWave mobile broadband (MMB): Unleashing vol. 18, pp. 4254–4267, Sep. 2019.
the 3–300 GHz spectrum,” in Proc. IEEE Sarnoff Symposium, May [109] H. Liu, H. Darabi, P. Banerjee, and J. Liu, “Survey of wireless indoor
2011, pp. 1–6. positioning techniques and systems,” IEEE Trans. Syst., Man, Cybern.
[87] A. F. Molisch, Wireless communications. John Wiley & Sons, 2012, C, vol. 37, no. 6, pp. 1067–1080, 2007.
vol. 34. [110] F. Zafari, A. Gkelias, and K. K. Leung, “A survey of indoor localization
[88] W. Roh, J.-Y. Seol, J. Park, B. Lee, J. Lee, Y. Kim, J. Cho, K. Cheun, systems and technologies,” IEEE Commun. Surveys Tuts., vol. 21, no. 3,
and F. Aryanfar, “Millimeter-wave beamforming as an enabling tech- pp. 2568–2599, 2019.
nology for 5G cellular communications: theoretical feasibility and [111] J. Xiao, Z. Zhou, Y. Yi, and L. M. Ni, “A survey on wireless indoor
prototype results,” IEEE Commun. Mag., vol. 52, pp. 106–113, Feb. localization from the device perspective,” ACM Comput. Surv., vol. 49,
2014. no. 2, pp. 1–31, 2016.
[89] S. Wane, D. Bajon, P. Corrales, M. Haider, J. Russer, Q.-H. Tran, C.-J. [112] R. M. Buehrer, H. Wymeersch, and R. M. Vaghefi, “Collaborative
Lin, S.-W. Chang, W.-T. Tsai, R. Giacometti et al., “Cognitive beam- sensor network localization: Algorithms and practical issues,” Proc.
former chips with smart-antennas for 5G and beyond: Holistic RFSOI IEEE, vol. 106, no. 6, pp. 1089–1114, 2018.
technology solutions including ASIC correlators,” in Proc. EuRAD, [113] F. Morselli, S. Bartoletti, M. Z. Win, and A. Conti, “Localization in
2019, pp. 453–456. 5G ecosystem with Wi-Fi,” in Proc. IEEE SPAWC, 2021, pp. 441–445.
[90] G. Menon, R. McMorrow, and D. Corman, “Active antennas for [114] A. Shahmansoori, G. E. Garcia, G. Destino, G. Seco-Granados, and
emerging x-band RADAR applications,” in Proc. IEEE RadarConf, H. Wymeersch, “5G position and orientation estimation through mil-
2018, pp. 1334–1337. limeter wave MIMO,” in Proc. IEEE GLOBECOM Wkshps, 2015, pp.
[91] E. Öjefors, M. Andreasson, T. Kjellberg, H. Berg, L. Aspemyr, 1–6.
R. Nilsson, K. Brink, R. Dahlbäck, D. Wu, K. Sjögren et al., “A 57-71 [115] R. Schmidt, “Multiple emitter location and signal parameter estima-
GHz beamforming SiGe transceiver for 802.11ad-based fixed wireless tion,” IEEE Trans. Antennas Propag., vol. 34, no. 3, pp. 276–280,
access,” in Proc. IEEE RFIC, 2018, pp. 276–279. 1986.
[92] E. L. Bengtsson, “Massive mimo from a terminal perspective,” Ph.D. [116] A. Paulraj, R. Roy, and T. Kailath, “Estimation of signal parameters
dissertation, PhD thesis, Lund University, 2019. via rotational invariance techniques – ESPRIT,” in Proc. Asilomar conf.
[93] H. Tataria, S. Sangodoyin, A. F. Molisch, P. J. Smith, M. Matthaiou, on signals syst. and comput., 1985, pp. 83–89.
J. Zhang, and R. S. Thomä, “Channel correlation diversity in MU- [117] Z. Yang, Z. Zhou, and Y. Liu, “From RSSI to CSI: Indoor localization
MIMO systems–analysis and measurements,” in Proc. IEEE PIMRC, via channel response,” ACM Comput. Surv., vol. 46, no. 2, pp. 1–32,
2019, pp. 1–7. Dec. 2013.
 40

[118] H. El-Sayed, G. Athanasiou, and C. Fischione, “Evaluation of local- [142] T. T. Tsai, L. H. Shen, C. J. Chiu, and K. T. Feng, “Beam AoD-based
ization methods in millimeter-wave wireless systems,” in Proc. IEEE Indoor Positioning for 60 GHz mmWave System,” in Proc. IEEE VTC-
CAMAD, 2014, pp. 345–349. Fall, 2020, pp. 1–5.
[119] J. Palacios, D. Steinmetzer, A. Loch, M. Hollick, and J. Widmer, [143] I. Pefkianakis and K.-H. Kim, “Accurate 3D localization for 60 GHz
“Adaptive codebook optimization for beam training on off-the-shelf networks,” in Proc. ACM SenSys, 2018, p. 120–131.
IEEE 802.11ad devices,” in Proc. ACM MobiCom, Oct. 2018, p. [144] G. Bielsa, J. Palacios, A. Loch, D. Steinmetzer, P. Casari, and J. Wid-
241–255. mer, “Indoor localization using commercial off-the-shelf 60 GHz access
[120] Y. Xie, J. Xiong, M. Li, and K. Jamieson, “MD-Track: Leveraging points,” in Proc. IEEE INFOCOM, 2018, pp. 2384–2392.
multi-dimensionality for passive indoor Wi-Fi tracking,” in Proc. ACM [145] O. Kanhere, S. Ju, Y. Xing, and T. S. Rappaport, “Map-assisted mil-
MobiCom, Aug. 2019, pp. 1–16. limeter wave localization for accurate position location,” in Proc. IEEE
[121] J. Palacios, G. Bielsa, P. Casari, and J. Widmer, “Single-and multiple- GLOBECOM, 2019, pp. 1–6.
access point indoor localization for millimeter-wave networks,” IEEE [146] N. Maletic, V. Sark, J. Gutiérrez, and E. Grass, “Device localization
Trans. Wireless Commun., vol. 18, no. 3, pp. 1927–1942, 2019. using mmWave ranging with sub-6-assisted angle of arrival estimation,”
[122] A. Olivier, G. Bielsa, I. Tejado, M. Zorzi, J. Widmer, and P. Casari, in Proc. IEEE BMSB, 2018, pp. 1–6.
“Lightweight indoor localization for 60-GHz millimeter wave systems,” [147] Z. Lin, T. Lv, and P. T. Mathiopoulos, “3-D indoor positioning for
in Proc. IEEE SECON, 2016, pp. 1–9. millimeter-wave massive MIMO systems,” IEEE Trans. Commun.,
[123] A. Fascista, A. Coluccia, H. Wymeersch, and G. Seco-Granados, “Low- vol. 66, no. 6, pp. 2472–2486, Jun. 2018.
complexity accurate mmwave positioning for single-antenna users [148] J. Chen, D. Steinmetzer, J. Classen, E. Knightly, and M. Hollick,
based on angle-of-departure and adaptive beamforming,” in Proc. IEEE “Pseudo lateration: Millimeter-wave localization using a single RF
ICASSP, 2020, pp. 4866–4870. chain,” in Proc. IEEE WCNC, 2017, pp. 1–6.
[124] M. Z. Comiter, M. B. Crouse, and H. T. Kung, “A data-driven [149] Y. Jia, H. Tian, S. Fan, and B. Liu, “Motion feature and millimeter
approach to localization for high frequency wireless mobile networks,” wave multi-path AoA-ToA based 3D indoor positioning,” in Proc. IEEE
in Proc. IEEE GLOBECOM, 2017, pp. 1–7. PIMRC, 2018, pp. 1–7.
[125] A. Yassin, Y. Nasser, M. Awad, and A. Al-Dubai, “Simultaneous [150] B. Hu, H. Tian, and S. Fan, “Millimeter wave LOS/NLOS identification
context inference and mapping using mm-Wave for indoor scenarios,” and localization via mean-shift clustering,” in Proc. IEEE PIMRC,
in Proc. IEEE ICC, 2017, pp. 1–6. 2019, pp. 1–7.
[126] J. Palacios, P. Casari, and J. Widmer, “JADE: Zero-knowledge device [151] F. Fellhauer, J. Lassen, A. Jaber, N. Loghin, and S. Ten Brink, “Non-
localization and environment mapping for millimeter wave systems,” line-of-sight positioning for mmwave communications,” in Proc. IEEE
in Proc. IEEE INFOCOM, 2017, pp. 1–9. SPAWC, 2018, pp. 1–5.
[127] J. Palacios, G. Bielsa, P. Casari, and J. Widmer, “Communication- [152] A. Zhou, S. Yang, Y. Yang, Y. Fan, and H. Ma, “Autonomous environ-
driven localization and mapping for millimeter wave networks,” in ment mapping using commodity millimeter-wave network device,” in
Proc. IEEE INFOCOM, 2018, pp. 2402–2410. Proc. IEEE INFOCOM, 2019, pp. 1126–1134.
[128] J. Palacios, P. Casari, H. Assasa, and J. Widmer, “LEAP: Location
[153] D. Garcia, J. O. Lacruz, P. Jiménez Mateo, and J. Widmer, “POLAR:
estimation and predictive handover with consumer-grade mmWave
Passive object localization with IEEE 802.11ad using phased antenna
devices,” in Proc. IEEE INFOCOM, 2019, pp. 2377–2385.
arrays,” in Proc. IEEE INFOCOM, 2020, pp. 1838–1847.
[129] Y. Lin, S. Jin, M. Matthaiou, and X. You, “Channel estimation and
[154] A. Yassin, Y. Nasser, and M. Awad, “Geometric approach in simulta-
indoor positioning for wideband multiuser millimeter wave systems,”
neous context inference, localization and mapping using mm-Wave,”
in Proc. IEEE SAM, 2020, pp. 1–5.
in Proc. ICT, 2018, pp. 159–164.
[130] H. Ajorloo, C. J. Sreenan, A. Loch, and J. Widmer, “On the feasibility
of using IEEE 802.11ad mmWave for accurate object detection,” in [155] A. Yassin, Y. Nasser, A. Y. Al-Dubai, and M. Awad, “MOSAIC:
Proc. ACM/SIGAPP SAC, 2019, p. 2406–2413. Simultaneous localization and environment mapping using mmWave
[131] “Talon AD7200,” 2020, accessed: Sept. 2021. [Online]. Available: without a-priori knowledge,” IEEE Access, vol. 6, pp. 68 932–68 947,
https://www.tp-link.com/us/home-networking/wifi-router/ad7200/ 2018.
[132] M. Vari and D. Cassioli, “mmWaves RSSI indoor network localization,” [156] A. Yassin, Y. Nasser, Y. Corre, G. Gougeon, and Y. Lostanlen, “3D
in Proc. IEEE ICC, 2014, pp. 127–132. localization and mapping using mm-Wave: What are the opportunities
[133] M. Pajovic, P. Wang, T. Koike-Akino, H. Sun, and P. V. Orlik, in vehicular and indoor environments?” in Proc. ICT, 2018, pp. 324–
“Fingerprinting-based indoor localization with commercial mmWave 330.
WiFi - part I: RSS and beam indices,” in Proc. IEEE GLOBECOM, [157] A. Shastri, J. Palacios, and P. Casari, “Millimeter wave localization
2019, pp. 1–6. with imperfect training data using shallow neural networks,” 2020,
[134] P. Hong, C. Li, H. Chang, Y. Hsueh, and K. Wang, “WBF-PS: arXiv preprint 2112.05008, accessed: Jan. 2022. [Online]. Available:
WiGig beam fingerprinting for UAV positioning system in GPS-denied https://arxiv.org/abs/2112.05008
environments,” in Proc. IEEE INFOCOM, 2020, pp. 1778–1787. [158] M. Petri and M. Ehrig, “A SoC-based SDR platform for ultra-high
[135] O. AlHory, O. Shoushara, H. AlSuri, M. ALShunnaq, and F. Awad, data rate broadband communication, radar and localization systems,”
“5G mmWave indoor location identification using beamforming and in Proc. Wireless Days, 2019, pp. 1–4.
RSSI,” in Proc. ICICS, 2020, pp. 91–95. [159] S. Saha, Y. Ghasempour, M. Haider, T. Siddiqui, P. Melo, N. Somanchi,
[136] A. Vashist, D. R. Bhanushali, R. Relyea, C. Hochgraf, A. Ganguly, P. D. L. Zakrajsek, A. Singh, R. Shyamsunder, O. Torres, D. Uvaydov,
Sai Manoj, R. Ptucha, A. Kwasinski, and M. E. Kuhl, “Indoor wireless J. Jornet, E. Knightly, D. Koutsonikolas, D. Pados, Z. Sun, and
localization using consumer-grade 60 GHz equipment with machine N. Thawdar, “X60: A programmable testbed for wideband 60 GHz
learning for intelligent material handling,” in Proc. IEEE ICCE, 2020, WLANs with phased arrays,” Computer Commun., vol. 133, 09 2018.
pp. 1–6. [160] S. Aggarwal, U. S. Sardesai, V. Sinha, and D. Koutsonikolas, “An
[137] P. Wang, M. Pajovic, T. Koike-Akino, H. Sun, and P. V. Orlik, experimental study of rate and beam adaptation in 60 GHz WLANs,”
“Fingerprinting-based indoor localization with commercial mmWave in Proc. ACM MSWiM, 2020, p. 171–180.
WiFi - Part II: Spatial beam SNRs,” in Proc. IEEE GLOBECOM, 2019, [161] J. Zhang, X. Zhang, P. Kulkarni, and P. Ramanathan, “OpenMili: a
pp. 1–6. 60 GHz software radio platform with a reconfigurable phased-array
[138] T. Koike-Akino, P. Wang, M. Pajovic, H. Sun, and P. V. Orlik, antenna,” in Proc. ACM MobiCom, 2016, pp. 162–175.
“Fingerprinting-based indoor localization with commercial MMWave [162] “Pi-radio,” 2020, accessed: Sept. 2021. [Online]. Available: https:
WiFi: A deep learning approach,” IEEE Access, vol. 8, pp. 84 879– //www.pi-rad.io
84 892, 2020. [163] “Zync ZCU111,” 2020, accessed: July 2021. [Online]. Available:
[139] P. Wang, T. Koike-Akino, and P. V. Orlik, “Fingerprinting-based indoor https://www.xilinx.com/products/boards-and-kits/zcu111.html
localization with commercial mmWave WiFi: NLOS propagation,” in [164] J. O. Lacruz, D. Garcia, P. J. Mateo, J. Palacios, and J. Widmer, “mm-
Proc. IEEE GLOBECOM, 2020, pp. 1–6. FLEX: an open platform for millimeter-wave mobile full-bandwidth
[140] N. Maletic, V. Sark, M. Ehrig, J. Gutiérrez, and E. Grass, “Experimental experimentation,” in Proc. ACM MobiSys, Toronto Ontario Canada,
evaluation of round-trip ToF-based localization in the 60 GHz band,” Jun. 2020, pp. 1–13.
in Proc. IPIN, 2019, pp. 1–6. [165] M. Polese, F. Restuccia, A. Gosain, J. Jornet, S. Bhardwaj, V. Ari-
[141] Z. Wei, Y. Zhao, X. Liu, and Z. Feng, “DoA-LF: A location fingerprint yarathna, S. Mandal, K. Zheng, A. Dhananjay, M. Mezzavilla et al.,
positioning algorithm with millimeter-wave,” IEEE Access, vol. 5, pp. “MillimeTera: Toward a large-scale open-source mmWave and terahertz
22 678–22 688, 2017. experimental testbed,” in Proc. ACM mmNets, 2019, pp. 27–32.
 41

[166] O. Kanhere and T. S. Rappaport, “Position locationing for millimeter [192] C. Wang, K. Yang, and X. Sun, “Precise localization of concealed
wave systems,” in Proc. IEEE GLOBECOM, 2018, pp. 206–212. objects in millimeter-wave images via semantic segmentation,” IEEE
[167] S. Wang, J. Huang, X. Zhang, H. Kim, and S. Dey, “X-Array: approx- Access, vol. 8, pp. 121 246–121 256, 2020.
imating omnidirectional millimeter-wave coverage using an array of [193] T. K. Vodai, K. Oleksak, T. Kvelashvili, F. Foroughian, C. Bauder,
phased arrays,” in Proc. ACM MobiCom, 2020, pp. 1–14. P. Theilmann, A. Fathy, and O. Kilic, “Enhancement of remote vital
[168] “Airfide Inc.” 2019, accessed: June 2021. [Online]. Available: sign monitoring detection accuracy using multiple-input multiple-
http://airfidenet.com output 77 GHz FMCW radar,” IEEE J. Electromagn., RF Microw. Med.
[169] D. Steinmetzer, D. Wegemer, M. Schulz, J. Widmer, and M. Hollick, Biol.,, pp. 1–1, 2021, in press.
“Compressive millimeter-wave sector selection in off-the-shelf IEEE [194] C. X. Lu, S. Rosa, P. Zhao, B. Wang, C. Chen, J. A. Stankovic,
802.11ad devices,” in Proc. ACM CoNEXT, 2017. N. Trigoni, and A. Markham, “See through smoke: robust indoor
[170] C. R. C. M. Da Silva, J. Kosloff, C. Chen, A. Lomayev, and mapping with low-cost mmWave radar,” in Proc. ACM MobiSys, 2020,
C. Cordeiro, “Beamforming training for IEEE 802.11ay millimeter pp. 14–27.
wave systems,” in Proc. IEEE ITA, 2018, pp. 1–9. [195] M. Jankiraman, FMCW Radar Design. Artech House, 2018.
[171] T. S. Rappaport, Y. Xing, O. Kanhere, S. Ju, A. Madanayake, S. Man- [196] C. Iovescu and S. Rao, “The fundamentals of millimeter wave sensors,”
dal, A. Alkhateeb, and G. C. Trichopoulos, “Wireless communications Texas Instruments, Tech. Rep. SPYY005A, 2017.
and applications above 100 GHz: Opportunities and challenges for 6G [197] A. E. Omer, S. Safavi-Naeini, R. Hughson, and G. Shaker, “Blood glu-
and beyond,” IEEE Access, vol. 7, pp. 78 729–78 757, 2019. cose level monitoring using an FMCW millimeter-wave radar sensor,”
Remote Sensing, vol. 12, no. 3, pp. 1–25, 2020.
[172] S.-H. Chen and S.-K. Jeng, “An SBR/image approach for radio wave
[198] A. E. Omer, G. Shaker, S. Safavi-Naeini, K. Murray, and R. Hughson,
propagation in indoor environments with metallic furniture,” IEEE
“Glucose levels detection using mm-wave radar,” IEEE Sensors Lett.,
Trans. Antennas Propag., vol. 45, no. 1, pp. 98–106, 1997.
vol. 2, no. 3, pp. 1–4, 2018.
[173] O. Kanhere and T. S. Rappaport, “Millimeter wave position location [199] T. Gu, Z. Fang, Z. Yang, P. Hu, and P. Mohapatra, “Mmsense: Multi-
using multipath differentiation for 3GPP using field measurements,” in person detection and identification via mmwave sensing,” in Proc. ACM
Proc. IEEE GLOBECOM, 2020, pp. 1–7. mmNets, 2019, pp. 45–50.
[174] ——, “Outdoor sub-THz position location and tracking using field [200] Z. Yang, P. H. Pathak, Y. Zeng, X. Liran, and P. Mohapatra, “Vital sign
measurements at 142 GHz,” in Proic. IEEE ICC, 2021, pp. 1–6. and sleep monitoring using millimeter wave,” ACM Trans. on Sensor
[175] “S 5GCHANNEL,” 2018, accessed: Dec 2021. Networks, vol. 13, no. 2, pp. 1–32, 2017.
[Online]. Available: https://www.siradel.com/s [201] J. Pegoraro, J. O. Lacruz, E. Bashirov, M. Rossi, and J. Widmer,
5gchannel-for-advanced-5g-mmw-propagation-modeling/ “RAPID: Retrofitting IEEE 802.11ay access points for indoor human
[176] F. Gu, X. Hu, M. Ramezani, D. Acharya, K. Khoshelham, S. Valaee, detection and sensing,” arXiv preprint arXiv:2109.04819, 2021.
and J. Shang, “Indoor localization improved by spatial context–a [202] M. Alizadeh, G. Shaker, and S. Safavi-Naeini, “Remote health moni-
survey,” ACM Comput. Surv., vol. 52, no. 3, pp. 1–35, Jul. 2019. toring system for bedbound patients,” in Proc. IEEE BIBE, 2020, pp.
[177] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image 801–806.
recognition,” in Proc. IEEE CVPR, 2016, pp. 770–778. [203] J. W. Smith, S. Thiagarajan, R. Willis, Y. Makris, and M. Torlak,
[178] O. Kanhere and T. S. Rappaport, “Position location for futuristic “Improved static hand gesture classification on deep convolutional
cellular communications: 5G and beyond,” IEEE Commun. Mag., neural networks using novel sterile training technique,” IEEE Access,
vol. 59, no. 1, pp. 70–75, 2021. vol. 9, pp. 10 893–10 902, 2021.
[179] C. Fiandrino, H. Assasa, P. Casari, and J. Widmer, “Scaling millimeter- [204] R. Zhang and S. Cao, “Real-time human motion behavior detection via
wave networks to dense deployments and dynamic environments,” CNN using mmWave radar,” IEEE Sensors J., vol. 3, no. 2, pp. 1–4,
Proc. IEEE, vol. 107, no. 4, pp. 732–745, 2019. 2018.
[180] A. Davoli, G. Guerzoni, and G. M. Vitetta, “Machine learning and deep [205] G. Tiwari and S. Gupta, “An mmWave radar based real-time contactless
learning techniques for colocated MIMO radars: A tutorial overview,” fitness tracker using deep CNNs,” IEEE Sensors J., vol. 21, no. 15, pp.
IEEE Access, vol. 9, pp. 33 704–33 755, 2021. 17 262–17 270, 2021.
[181] B. van Berlo, A. Elkelany, T. Ozcelebi, and N. Meratnia, “Millimeter [206] F. Jin, R. Zhang, A. Sengupta, S. Cao, S. Hariri, N. K. Agarwal, and
wave sensing: A review of application pipelines and building blocks,” S. K. Agarwal, “Multiple patients behavior detection in real-time using
IEEE Sensors J., vol. 21, no. 9, pp. 10 332–10 368, 2021. mmWave radar and deep CNNs,” in Proc. IEEE RadarConf, 2019, pp.
[182] Y. Wang, H. Liu, K. Cui, A. Zhou, W. Li, and H. Ma, “m-Activity: 1–6.
Accurate and real-time human activity recognition via millimeter wave [207] A. Sengupta, F. Jin, R. Zhang, and S. Cao, “Mm-pose: Real-time human
radar,” in Proc. IEEE ICASSP, 2021, pp. 8298–8302. skeletal posture estimation using mmWave radars and CNNs,” IEEE
[183] J. Pegoraro, F. Meneghello, and M. Rossi, “Multiperson continuous Sensors J., vol. 20, no. 17, pp. 10 032–10 044, 2020.
tracking and identification from mm-Wave micro-doppler signatures,” [208] D. F. Albuquerque, E. S. Goncalves, E. F. Pedrosa, F. C. Teixeira, and
IEEE Trans. Geosci. Remote Sens., vol. 59, no. 4, pp. 2994–3009, 2021. J. N. Vieira, “Robot self position based on asynchronous millimetre
[184] J. Pegoraro and M. Rossi, “Real-time people tracking and identification wave radar interference,” in Proc. IPIN, 2019.
from sparse mm-wave radar point-clouds,” IEEE Access, vol. 9, pp. [209] F. Jin, A. Sengupta, and S. Cao, “mmFall: Fall detection using 4-
78 504–78 520, 2021. D mmWave radar and a hybrid variational RNN autoencoder,” IEEE
Trans. Autom. Sci. Eng., pp. 1–13, 2020, in press.
[185] G. Li, Z. Zhang, H. Yang, J. Pan, D. Chen, and J. Zhang, “Capturing
[210] H. Cui and N. Dahnoun, “High precision human detection and tracking
human pose using mmWave radar,” in Proc. IEEE PerCom Workshops,
using millimeter-wave radars,” IEEE Aerosp. Electron. Syst. Mag.,
2020, pp. 1–6.
vol. 36, no. 1, pp. 22–32, 2021.
[186] C. Jiang, J. Guo, Y. He, M. Jin, S. Li, and Y. Liu, “Mmvib: Micrometer- [211] J. Wu, H. Cui, and N. Dahnoun, “A novel high performance human
level vibration measurement with mmwave radar,” in Proc. ACM detection, tracking and alarm system based on millimeter-wave radar,”
MobiCom, 2020, pp. 1–13. in Proc. MECO, 2021, pp. 1–4.
[187] S. D. Regani, C. Wu, B. Wang, M. Wu, and K. J. R. Liu, “mmWrite: [212] V. C. Chen, F. Li, S.-S. Ho, and H. Wechsler, “Micro-doppler effect in
Passive handwriting tracking using a single millimeter wave radio,” radar: phenomenon, model, and simulation study,” IEEE Trans. Aerosp.
IEEE Internet Things J., vol. 8, no. 17, pp. 13 291–13 305, Sep. 2021. Electron. Syst., vol. 42, no. 1, pp. 2–21, 2006.
[188] D. Nickalls, J. Wu, and N. Dahnoun, “A real-time and high perfor- [213] J. B. Allen and L. R. Rabiner, “A unified approach to short-time fourier
mance posture estimation system based on millimeter-wave radar,” in analysis and synthesis,” Proc. IEEE, vol. 65, no. 11, pp. 1558–1564,
Proc. MECO, 2021, pp. 1–4. 1977.
[189] M. Z. Ozturk, C. Wu, B. Wang, and K. J. R. Liu, “GaitCube: Deep [214] L. Cohen, Time-frequency analysis. Prentice Hall, 1995, vol. 778.
data cube learning for human recognition with millimeter-wave radio,” [215] V. Chen, D. Tahmoush, and W. Miceli, Radar Micro-Doppler Signa-
IEEE Internet Things J., vol. 9, no. 1, pp. 546–557, Jan. 2022. tures: Processing and Applications. IET, 2014.
[190] P. Zhao, C. X. Lu, B. Wang, C. Chen, L. Xie, M. Wang, N. Trigoni, [216] R. E. Kalman, “A New Approach to Linear Filtering and Prediction
and A. Markham, “Heart rate sensing with a robot mounted mmwave Problems,” J. Basic Eng., vol. 82, no. 1, pp. 35–45, 1960.
radar,” in Proc. IEEE ICRA, 2020, pp. 2812–2818. [217] G. Welch, G. Bishop et al., “An introduction to the kalman filter,”
[191] H. Cui and N. Dahnoun, “High precision human detection and tracking University of North Carolina at Chapel Hill, Dept. of Computer
using millimeter-wave radars,” IEEE Aerosp. Electron. Syst. Mag., Science, Tech. Rep. TR 95-041, 1995, accessed: Sept. 2021. [Online].
vol. 36, pp. 22–32, 2021. Available: https://www.cs.unc.edu/∼welch/media/pdf/kalman intro.pdf
 42

[218] S. J. Julier and J. K. Uhlmann, “Reduced sigma point filters for the [244] B. Kapilevich, Y. Pinhasi, M. Anisimov, B. Litvak, and D. Hardon,
propagation of means and covariances through nonlinear transforma- “FMCW mm-wave non-imaging sensor for detecting hidden objects,”
tions,” in Proc. IEEE ACC, vol. 2, 2002, pp. 887–892. in Proc. IEEE MTT-S IMWS, 2011, pp. 101–104.
[219] E. A. Wan and R. Van Der Merwe, “The unscented Kalman filter for [245] B. Kapilevich, B. Litvak, and A. Shulzinger, “Passive non-imaging mm-
nonlinear estimation,” in Proc. IEEE AS-SPCC, 2000, pp. 153–158. wave sensor for detecting hidden objects,” in Proc. IEEE COMCAS,
[220] T. M. Mitchell, Machine learning. McGraw-Hill, 1997. 2013, pp. 1–5.
[221] I. Goodfellow, Y. Bengio, A. Courville, and Y. Bengio, Deep learning. [246] J. Bhatia, A. Dayal, A. Jha, S. K. Vishvakarma, J. Soumya, M. B.
MIT press Cambridge, 2016, vol. 1, no. 2. Srinivas, P. K. Yalavarthy, A. Kumar, V. Lalitha, S. Koorapati, and L. R.
[222] M. Ester, H.-P. Kriegel, J. Sander, X. Xu et al., “A density-based Cenkeramaddi, “Object classification technique for mmWave FMCW
algorithm for discovering clusters in large spatial databases with noise,” radars using range-FFT features,” in Proc. COMSNETS, 2021, pp. 111–
in Proc. AAAI KDD, vol. 96, no. 34, 1996, pp. 226–231. 115.
[223] V. Estivill-Castro, “Why so many clustering algorithms: a position [247] S. Bakhtiari, S. Liao, T. Elmer, A. Raptis et al., “A real-time heart rate
paper,” ACM SIGKDD explorations newsletter, vol. 4, no. 1, pp. 65–75, analysis for a remote millimeter wave IQ sensor,” IEEE Trans. Biomed.
2002. Eng., vol. 58, no. 6, pp. 1839–1845, 2011.
[248] J. Lien, N. Gillian, M. E. Karagozler, P. Amihood, C. Schwesig,
[224] D. Kellner, J. Klappstein, and K. Dietmayer, “Grid-based DBSCAN for
E. Olson, H. Raja, and I. Poupyrev, “Soli: Ubiquitous gesture sensing
clustering extended objects in radar data,” in Proc. IEEE IVS, 2012,
with millimeter wave radar,” ACM Trans. on Graphics, vol. 35, no. 4,
pp. 365–370.
pp. 1–19, 2016.
[225] U. Hasson, S. A. Nastase, and A. Goldstein, “Direct fit to nature: An [249] C. Xu, Z. Li, H. Zhang, A. S. Rathore, H. Li, C. Song, K. Wang, and
evolutionary perspective on biological and artificial neural networks,” W. Xu, “Waveear: Exploring a mmwave-based noise-resistant speech
Neuron, vol. 105, no. 3, pp. 416–434, 2020. sensing for voice-user interface,” in Proc. ACM MobiSys, 2019, pp.
[226] F. Rosenblatt, “The perceptron: a probabilistic model for information 14–26.
storage and organization in the brain.” Psychological review, vol. 65, [250] T.-Y. J. Kao, Y. Yan, T.-M. Shen, A. Y.-K. Chen, and J. Lin, “Design
no. 6, p. 386, 1958. and analysis of a 60-GHz CMOS doppler micro-radar system-in-
[227] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based package for vital-sign and vibration detection,” IEEE Trans. Microw.
learning applied to document recognition,” Proc. IEEE, vol. 86, no. 11, Theory Tech., vol. 61, no. 4, pp. 1649–1659, 2013.
pp. 2278–2324, 1998. [251] “TI mmWave radar sensors,” accessed: Mar. 2022. [Online]. Available:
[228] S. Chang, Y. Zhang, F. Zhang, X. Zhao, S. Huang, Z. Feng, and Z. Wei, https://www.ti.com/sensors/mmwave-radar/overview.html
“Spatial attention fusion for obstacle detection using mmwave radar and [252] “MMWCAS-RF-EVM,” accessed: Mar. 2022. [Online]. Available:
vision sensor,” Sensors, vol. 20, no. 4, pp. 1–21, 2020. https://www.ti.com/tool/MMWCAS-RF-EVM
[229] R. J. Williams and J. Peng, “An efficient gradient-based algorithm for
on-line training of recurrent network trajectories,” Neural computation,
vol. 2, no. 4, pp. 490–501, 1990.
[230] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[231] S. Hochreiter, “The vanishing gradient problem during learning re-
current neural nets and problem solutions,” International Journal of
Uncertainty, Fuzziness and Knowledge-Based Systems, vol. 6, no. 02,
pp. 107–116, 1998. Anish Shastri (S’18) is an Early Stage Researcher
[232] P. Zhao, C. X. Lu, J. Wang, C. Chen, W. Wang, N. Trigoni, and with the EU H2020 Marie Skłodowska Curie Ac-
A. Markham, “mID: Tracking and identifying people with millimeter tions MINTS ETN, pursuing his Ph.D. at the De-
wave radar,” in Proc. DCOSS, 2019, pp. 33–40. partment of Information Engineering and Computer
[233] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol, Science, University of Trento, Italy. He received his
“Stacked denoising autoencoders: Learning useful representations in a B.Tech in Electronics and Communication Engineer-
deep network with a local denoising criterion,” J. Mach. Learn. Res., ing (ECE) from the National Institute of Technology
vol. 11, p. 3371–3408, 2010. (MANIT) - Bhopal, India, in 2016, and MS by
[234] S. Wagner and W. Johannes, “Target detection using autoencoders in a Research in ECE from the International Institute
radar surveillance system,” in Proc. IEEE RadarConf, 2019, pp. 1–5. of Information Technology – Hyderabad (IIIT-H),
[235] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, India, in 2019. Before starting his Ph.D., he was a
S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial net- Research Intern in the Indoor Networks Research group at Nokia Bell Labs,
works,” Commun. ACM, vol. 63, no. 11, p. 139–144, Oct. 2020. Ireland. His research interests include signal processing and machine learning
[236] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep for wireless communications, with his current research focusing on algorithms
network training by reducing internal covariate shift,” in Proc. ICML, for mmWave indoor localization, environment mapping, and sensing.
2015, pp. 448–456.
[237] C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning
on point sets for 3D classification and segmentation,” in Proc. IEEE
CVPR, 2017, pp. 652–660.
[238] C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical
feature learning on point sets in a metric space,” 2017, arXiv preprint
1706.02413.
[239] W. Wu, Z. Qi, and L. Fuxin, “PointConv: Deep convolutional networks
on 3D point clouds,” in Proc. IEEE CVPR, 2019, pp. 9613–9622. Neharika Valecha (S’20) is an Early Stage Re-
[240] Z. Meng, S. Fu, J. Yan, H. Liang, A. Zhou, S. Zhu, H. Ma, J. Liu, and searcher with the EU H2020 Marie Skłodowska
N. Yang, “Gait recognition for co-existing multiple people using mil- Curie Actions MINTS ETN, pursuing her Ph.D.
limeter wave sensing,” in Proc. AAAI Conf. on Artificial Intelligence, at the Department of Electrical and Information
2020, pp. 849–856. Technology, Lund University, Sweden. She received
[241] D.-S. Lee, S. Yeom, J.-Y. Son, and S.-H. Kim, “Automatic image her Masters in Mobile Computing Systems from
segmentation for concealed object detection using the expectation- Institut Eurécom in 2019. Her research interests
maximization algorithm,” Optics express, vol. 18, no. 10, pp. 10 659– include signal processing and communication for
10 667, 2010. next generation wireless systems. Currently, she is
[242] S. Yeom, D.-S. Lee, J.-Y. Son, M.-K. Jung, Y. Jang, S.-W. Jung, and working with sensing and localization for mmWave
S.-J. Lee, “Real-time outdoor concealed-object detection with passive indoor scenarios.
millimeter wave imaging,” Optics express, vol. 19, no. 3, pp. 2530–
2536, 2011.
[243] T. Wei and X. Zhang, “Mtrack: High-precision passive tracking using
millimeter wave radios,” in Proc. ACM MobiCom, 2015, p. 117–129.
 43

Enver Bashirov (S’20) is currently an Early Stage Fredrik Tufvesson (F’17) received his Ph.D. degree
Researcher at EU Horizon 2020 Marie Skłodowska from Lund University, Lund, Sweden, in 2000. After
Curie project MINTS, pursuing his Ph.D. degree at two years at a startup company, he joined the De-
the Department of Information Engineering, Univer- partment of Electrical and Information Technology,
sity of Padova, Italy. He received his B.Sc. degree Lund University, where he is currently a Professor
in Computer Engineering from Bilkent University of radio systems. His main research interest is the
and M.Sc. degree in Applied Mathematics and Com- interplay between the radio channel and the rest of
puter Science from Eastern Mediterranean Univer- the communication system with various applications
sity, North Cyprus. His research interests include in 5G/B5G systems such as massive MIMO, mm
sensing applications in mmWave, together with ma- wave communication, vehicular communication and
chine learning and signal processing solutions. radio based positioning. Fredrik has authored around
100 journal papers and 150 conference papers, he is an IEEE Fellow, and his
research has been awarded with the Neal Shepherd Memorial Award (2015)
for the best propagation paper in the IEEE T RANSACTIONS ON V EHICULAR
T ECHNOLOGY, and with the IEEE Communications Society best tutorial
paper award (2018, 2021).

Michele Rossi (SM’13) is a Professor of Wireless

Networks in the Department of Information Engi-
Harsh Tataria (M’17) received the B.E. degree
neering (DEI) at the University of Padova (UNIPD),
(honors) in electronic and computer systems engi-
Italy, where is the head of the Master’s Degree in
neering and the Ph.D. degree in communications
ICT for internet and Multimedia (http://mime.dei.
engineering from the Victoria University of Welling-
unipd.it/). He also teaches Human Data Analysis at
ton, New Zealand, in December 2013 and March
the Data Science Master’s degree at the Department
2017, respectively. Since then, he has held post-
of Mathematics (DM) at UNIPD (https://datascience.
doctoral fellowship positions at Queen’s University
math.unipd.it/). Since 2017, he has been the Director
Belfast, Belfast, U.K., the University of Southern
of the DEI/IEEE Summer School of Information
California, Los Angeles, CA, USA, and Lund Uni-
Engineering (http://ssie.dei.unipd.it/). His research
versity, Sweden. His research interests include mea-
interests broadly embrace wireless sensing systems, green mobile networks,
surement and modeling of propagation channels,
edge and wearable computing. Over the years, he has been involved in several
multiple antenna transceiver design, and statistical analysis techniques of
EU projects on wireless sensing and IoT and has collaborated with major
multiple antenna systems at centimeter-wave, millimeter-wave, and sub-
companies such as Ericsson, DOCOMO, Samsung and Intel. His research
terahertz frequencies.
is currently supported by the European Commission through the H2020
projects MINTS (no. 861222) on “mmWave networking and sensing” and
GREENEDGE (no. 953775) on “green edge computing for mobile networks”
(project coordinator). Prof. Rossi has been the recipient of seven best paper
awards from the IEEE and currently serves on the Editorial Boards of the
IEEE T RANSACTIONS ON M OBILE C OMPUTING, and of the O PEN J OURNAL
OF THE C OMMUNICATIONS S OCIETY .

Michael Lentmaier (SM’11) is an Associate Profes- Paolo Casari received the PhD in Information En-
sor at the Department of Electrical and Information gineering in 2008 from the University of Padova,
Technology at Lund University, Sweden, which he Italy. He was on leave at the Massachusetts Institute
joined in January 2013. His research interests in- of Technology in 2007, working on underwater com-
clude design and analysis of coding systems, graph munications and networks. He collaborated to sev-
based iterative algorithms and Bayesian methods eral funded projects including EU FP7 and H2020
applied to decoding, detection and estimation in efforts, EDA projects, as well as US ARO, ONR and
communication systems. He received the Dipl.-Ing. NSF initiatives, and is currently the PI of the NATO
degree in electrical engineering from University of SPS project SAFE-UComm. In 2015, he joined the
Ulm, Germany in 1998, and the Ph.D. degree in IMDEA Networks Institute, Madrid, Spain, where
telecommunication theory from Lund University, in he led the Ubiquitous Wireless Networks group. In
2003. He then worked as a Post-Doctoral Research Associate at University October 2019, he joined the faculty of the University of Trento, Italy, as a
of Notre Dame, Indiana and at University of Ulm. From 2005 to 2007 tenure-tracked assistant professor.
he was with the Institute of Communications and Navigation of the Ger- Dr. Casari is currently on the editorial boards of the IEEE T RANSACTIONS
man Aerospace Center (DLR) in Oberpfaffenhofen, where he worked on ON M OBILE C OMPUTING and of the IEEE T RANSACTIONS ON W IRELESS
signal processing techniques in satellite navigation receivers. From 2008 C OMMUNICATIONS, and regularly serves in the organizing committee of
to 2012 he was a senior researcher and lecturer at the Vodafone Chair several international conferences. Previously, he has been guest editor of a
Mobile Communications Systems at TU Dresden, where he was heading special issue of IEEE ACCESS on “Underwater Acoustic Communications
the Algorithms and Coding research group. He is a senior member of the and Networking.” He received two best paper awards. His research interests
IEEE and served as an editor for the IEEE C OMMUNICATIONS L ETTERS include diverse aspects of networked communications and computing, such as
(2010-2013), IEEE T RANSACTIONS ON C OMMUNICATIONS (2014-2017), channel modeling, network protocol design, localization, resource allocation,
and IEEE T RANSACTIONS ON I NFORMATION T HEORY (2017-2020). He was simulation, and experimental evaluation.
awarded the Communications Society & Information Theory Society Joint
Paper Award (2012) for his paper “Iterative decoding threshold analysis for
LDPC convolutional codes.”