Step 1.

Formulate a clear task of my work

A survey and theoretical analysis of existing indoor positioning systems with current ML-based
techniques for their implementation in supermarkets.

Step 2. List of specific requirements (KPI):

1. Accuracy: This KPI measures the system's accuracy in determining the location of objects or
individuals within the indoor environment, expressed in meters. It reflects the precision and
reliability of the positioning system's output.

2. Coverage: This KPI assesses the extent of the indoor area covered by the positioning system,
indicating whether it provides coverage at the floor or room level. It ensures that the system can
track objects or individuals across the entire intended area of deployment without significant
blind spots.

3. Cost: This KPI evaluates the overall cost associated with implementing and maintaining the
indoor positioning system. It includes both initial setup costs (e.g., hardware, software,
installation) and ongoing expenses (e.g., maintenance, upgrades). Cost-effectiveness is crucial
for ensuring the feasibility and sustainability of the system deployment.

4. Energy: This KPI measures the energy efficiency of the indoor positioning system, reflecting
its power consumption and battery life. It assesses how effectively the system utilizes energy
resources to operate sensors, transmit data, and perform computations, thereby influencing its
long-term reliability and operational sustainability.

5. Computation: This KPI evaluates the computational requirements of the positioning system,
including processing power, memory usage, and data processing efficiency. It assesses the
system's ability to handle complex algorithms, real-time data processing, and large datasets,
which are essential for accurate and responsive indoor location tracking.

6. Pros: This KPI identifies the advantages and strengths of the indoor positioning system,
highlighting its key features, functionalities, and potential benefits. It provides a comprehensive
overview of the positive aspects that make the system attractive for deployment in various indoor
environments.

7. Cons: This KPI identifies the limitations and challenges associated with the indoor positioning
system, including potential drawbacks, limitations, and areas for improvement. It helps
stakeholders understand the weaknesses and constraints of the system, enabling informed
decision-making and risk management.
 Tasks:

1. Integrate Table III & Fig.2 to create a table with existing algorithms
2. Find info for the table of Accuracy to change words to real numbers. Same for
Complexity&Cost – how I estimate it??? Look at papers
3. Describe every technology from the list in a term of ML-based tools, note prons &
cons of implementing in supermarkets – work with texts, add theoretical things

4. Future predictions
5. Conclusion

https://www.researchgate.net/profile/Nur-Abdul-Wahab-4/publication/
361748336_Indoor_Positioning_System_A_Review/links/62d563bbef252b2e6ddf815d/Indoor-
Positioning-System-A-Review.pdf

https://ieeexplore.ieee.org/abstract/document/8683984

Technol Techniq Accuracy Coverag Cost Energy Comp Pros Cons

ogy ue (m) e utatio
n
Satellite- Trilateration 5-10m Floor level Low Low Low Low power • Poor indoor
consumption performance
Based(G
PS) • Signal Quality
• Highly Variability
ubiquitous

• Signal
• Wide Area Interference
Coverage
Magneti Trilateration 2 Floor level Low Low Medium Cheap Requires mapping
Fingerprintin
c-Based
g • No pre-deployed • Complexity
infrastructure is increases as the
required number of sensors
increases
• Magnetic field is
everywhere and • Accuracy
relatively stable depends on the
variation in
magnetic field.

Sound
Based
Audible
Sound
Ultrason Trilateration 0.01–1 Room level Mediu Medium High Good accuracy Interference
m– No effect of
ic
High multipath Cost for hardware

• NoneedofLoS • Sensitive to
• Low temperature and
Interference pressure
• Available on
Smartphones
• No penetration
though some
solids

Acoustic 1-5m Room level Low Medium High Cheap Poor accuracy

Sound Trilateration
• NoneedofLoS • Sensitive to
• Low environmental
Interference conditions
• Available on
Smartphones

Radio
Frequen
cy-
Based
Wi-Fi Proximity 1-5 Floor level Low High Medium Good accuracy RF interference
Trilateration (around 35) Low cost Wi-Fi with devices
Angulation signals can operating at 2.4
Fingerprintin penetrate walls/ GHz
g RSS- No need for Fingerprinting
Propagation additional requires a huge
model infrastructure effort

• Use existing • Highly

communication influenced by
networks environmental
changes.
• Majority of
devices available • Security
nowadays are concerns
equipped with
WLAN
connectivity

• LOS is not
required

BLE Proximity 2-5 Floor level Low- Low Low Good accuracy RF interference
Trilateration Mediu No need for
Fingerprintin m additional • Relatively
g infrastructure Low expensive
power
consumption • High radio
interference

• Do not require
LOS

• A lighter
standard
• Highly
ubiquitous

RFDI Proximity 1–5 Low Low Medium Cheap Real-time Low accuracy
Trilateration localization Response time is
Fingerprintin Room level high
g RSS- • LOS is not
Propagation required between • Less compatible
RF transmitters with other
model
and receivers technolo- gies

• Penetrate • The data signal

through solid and is affected by
non-metal objects antenna

• Security issues
UWB Trilateration 0.01–1 Few meters High Medium Medium Accurate Expensive
Angulation Coverage is
• Penetrate limited
through walls Performance
• No interference
degrades in NLOS
with existing RF
systems
• Interference due
to metalic and
liquid materials

Optical-
Based
IR Proximity 1-2 Room level Mediu Low Low Cheap No effect Sunlight
Trilateration (few meters) m of multipath interference
Short-range
Low power
consumption Cost for hardware

• Low Signal • Line-of-Sight

Interference Requirement
• Sensitivity to
Lighting
• Enhanced
Conditions
privacy
• Fine-Grained
Positioning • Limited Range

VLC Angulation 0.1 Floor level Mediu Medium Medium No interfering Expensive
m construction
• Supports larger
bandwidth • Interference
• Secured issues from other
communication light sources
• No interference
due to EM
radiations • Requires both
source and
receiver should be
• Easy to install in LOS

• Easily affected
by atmospheric
absorp- tion,
shadowing, and
beam dispersion
 1. Let’s select existing general IPS technologies through the lens of general
KPI(accuracy, coverage, cost):

High-med accur, coverage>40m, low-med cost

WLAN, VLC, Bluetooth, Magnetic

A. Visible Light Communication (VLC)-based

Visible light communication (VLC) using LEDs has garnered attention due to its energy
efficiency and cost-effectiveness, especially for indoor illumination, and its resilience to
electromagnetic interference. Consequently, research efforts have intensified to improve indoor
visible light positioning (VLP) systems. Photo-diodes (PDs) are preferred over image sensors for
VLC detection due to their affordability and simplicity, particularly with the RSS algorithm,
which is more reliable in larger scenarios [95]. However, RSS-based VLP suffers from
positioning errors, prompting researchers to explore ML-based methods to enhance accuracy.

Various ANN-based visible light indoor positioning systems are proposed in [92], [93] and [94].
In [92], an LED bulb is considered as the transmitter, and a photo diode is considered as the
receiver. The communication channel is modeled as a visible light channel influenced by
multipath effects which is removed by the application of ANN. The system used the Combined
Deterministic and Monte Carlo (CDMMC) method for localization. In [93], the system used a
back-propagation ANN with optical camera communications where LED lights are clustered into
blocks and these coordinates are encoded using under sampled modulation scheme. The location
of the camera receiver) is localized using the estimated coordinates of the blocks. Two indoor
positioning methods using VLS - based on ANN are proposed by Shencheng et al. [94]. In this
approach, the first method creates 4 networks that are used for training and estimating the
coordinates of the target location. The second method creates only one network with 4 input
layers. The system divides the entire indoor space into small blocks. The receiver captures the
signal sent by each LED with ID information with which the system localizes the receiver. The
transmitted optical signal is modulated using asymmetric clipped optical orthogonal frequency
division multiplexing (ACO-OFDM) technology. These 2 methods are robust and achieved a
mean position error of 3.29cm and 2.78cm, respectively.
The utilization of VLC combined with a genetic algorithm in indoor localization, as presented in
[95], involves calculating 3D coordinates within the optical wireless environment using
multipath reflections. A modified genetic algorithm facilitates global optimization for 3D
position estimation without prior knowledge of device height or orientation. ANN processes
first-order reflections, achieving a low average localization error of 1.02 cm. 3D visible light
positioning (3DVLP) proposed in [100] excels in accuracy and consistency. It employs a
regression neural network to estimate real-time target location. Image sensors capture AoA
information from transmitting LED lights for network input. The offline phase preprocesses the
data extracted from reference points, which trains the network using Adam optimizer. The
trained network in the online phase detects target locations with a mean error of 1.1 cm.

Sayed et al. introduced two VLC-based methods to locate a moving user in [113]. VLC
transmitters convey location info through visible light, which a photodetector carried by the
target receives. Trilateration and neural network approaches are used for instantaneous
prediction. For LOS, the system’s max positioning error is 2.9 cm; for NLoS, it’s 8.1 cm, with
the neural network outperforming trilateration. In [114], a Position Estimation Deep Neural
Network (PE-DNN) aided VLC system is introduced to address complexity and compatibility
issues. X. Lin’s approach employs a DNN for processing data, enabling 2D location estimation
with just one LED transmitter. Achieving centimeter-level accuracy, the system attains a
minimum positioning error of 4.18 cm. In [115], a trained neural network mitigates the impact of
indoor diffuse channels in VLC positioning. Utilizing RSS data and backpropagation, the
algorithm achieves an average positioning error of 6.59 cm.

VLC-based IPSs encounter challenges due to signal blockage, lighting conditions, and LoS
constraints. Complex algorithms for angle estimation and decoding, specialized hardware, and
integration with lighting systems affect cost and scalability. Addressing these is essential for
improving VLC-based IPS accuracy and feasibility. A comparison of VLC-based IPSs
considering factors like complexity, scalability, and cost is given in Table. IV.
Visible Light Communication (VLC) technology utilizes LEDs for communication, offering
energy efficiency, cost-effectiveness, and resilience to electromagnetic interference. In the
context of supermarkets, implementing VLC-based indoor positioning systems (IPS) can have
several advantages and disadvantages, especially when viewed through the lens of ML-based
tools.

**Pros:**

1. **Enhanced Accuracy**: ML-based methods, such as Artificial Neural Networks (ANN), can
improve the accuracy of VLC-based indoor positioning systems by processing received signal
strength (RSS) data and mitigating positioning errors caused by multipath effects. This increased
accuracy is crucial for tracking inventory, optimizing shelf layouts, and analyzing customer
behavior within the supermarket.

2. **Real-time Tracking**: ML algorithms facilitate real-time tracking of moving users within

the supermarket. This capability can be leveraged for personalized marketing strategies, efficient
queue management, and monitoring of high-traffic areas.

3. **Cost-effectiveness**: Utilizing photo-diodes (PDs) as receivers for VLC detection, coupled

with ML algorithms, offers a cost-effective solution for indoor positioning compared to complex
image sensor-based systems. This affordability is advantageous for widespread deployment
across the supermarket.

4. **Centimeter-level Accuracy**: ML-based VLC systems can achieve centimeter-level

accuracy in positioning, enabling precise localization of products and customers. This accuracy
is essential for tasks such as inventory management, asset tracking, and location-based
promotions.

5. **Adaptability**: ML algorithms can adapt to changing environmental conditions, such as

variations in lighting and signal blockage, ensuring reliable performance of VLC-based IPS in
diverse settings within the supermarket.
**Cons:**

1. **Complexity**: Implementing ML-based VLC systems requires expertise in both machine

learning and optical communication technologies. This complexity may increase the initial setup
and maintenance costs, as well as the need for specialized personnel to manage the system.

2. **Hardware Requirements**: While PDs are preferred for VLC detection due to their
affordability, ML-based algorithms may require specialized hardware for processing and
analyzing data. This additional hardware can contribute to the overall cost of implementing
VLC-based IPS in the supermarket.

3. **Integration Challenges**: Integrating VLC-based IPS with existing supermarket

infrastructure, such as lighting systems and checkout counters, may pose challenges. Ensuring
seamless integration without disrupting regular operations requires careful planning and
coordination.

4. **Scalability**: Scaling up VLC-based IPS to cover larger areas of the supermarket may
introduce scalability challenges, particularly in terms of managing data processing and
maintaining consistent performance across the entire space.

5. **Signal Interference**: VLC signals may face interference from ambient lighting conditions
and obstacles within the supermarket, leading to inaccuracies in positioning. ML algorithms can
mitigate some of these effects but may not completely eliminate them.

In conclusion, while ML-based VLC technology offers several benefits for indoor positioning in
supermarkets, including enhanced accuracy and real-time tracking, its implementation requires
careful consideration of factors such as complexity, hardware requirements, and integration
challenges. Addressing these concerns is essential for maximizing the effectiveness and
feasibility of VLC-based IPS in supermarket environments.
Certainly, let's delve deeper into how ML-based VLC technology impacts key performance
indicators (KPIs) such as accuracy, coverage, and cost & complexity within the context of
implementing indoor positioning systems (IPS) in supermarkets.

**Accuracy:**

- *Pros*: ML algorithms, especially those based on Artificial Neural Networks (ANN), can
significantly enhance the accuracy of VLC-based IPS by processing complex data patterns and
mitigating errors caused by multipath effects and signal interference. This increased accuracy
ensures precise localization of products, customers, and assets within the supermarket, thereby
improving inventory management, customer service, and targeted marketing efforts.

- *Cons*: While ML algorithms improve accuracy, there may still be limitations in certain
scenarios, such as areas with dense shelving or high levels of ambient light, which can impact
signal quality and accuracy. Additionally, achieving and maintaining high accuracy may require
continuous monitoring and recalibration of the system, adding to operational complexity.

**Coverage:**

- *Pros*: ML-based VLC systems can provide extensive coverage within the supermarket,
leveraging the widespread deployment of LED lighting fixtures for communication. This
coverage ensures that indoor positioning capabilities are available across various sections and
aisles, enabling comprehensive tracking of assets and customers throughout the retail space.

- *Cons*: Despite the potential for extensive coverage, there may be limitations in areas with
signal blockage or obstructions, such as refrigeration units, dense merchandise displays, or
architectural features. Addressing these coverage gaps may require additional infrastructure
investments or strategic placement of VLC transmitters, adding to the overall complexity and
cost of the system.

**Cost & Complexity:**

- *Pros*: ML-based VLC technology offers cost-effective solutions for indoor positioning in
supermarkets, leveraging affordable photo-diodes (PDs) for VLC detection and processing.
Additionally, ML algorithms can optimize the utilization of existing hardware and infrastructure,
minimizing the need for expensive upgrades or replacements.

- *Cons*: Despite the potential for cost savings, implementing ML-based VLC systems
introduces complexity in terms of system design, integration, and maintenance. The expertise
required for developing and managing ML algorithms, coupled with the need for specialized
hardware and software, can contribute to upfront costs and ongoing operational expenses.
Furthermore, integrating VLC-based IPS with existing supermarket infrastructure, such as
lighting systems and inventory management software, may require significant customization and
coordination, increasing both cost and complexity.

In summary, while ML-based VLC technology offers promising benefits for indoor positioning
in supermarkets, including improved accuracy, extensive coverage, and cost-effective solutions,
its implementation requires careful consideration of factors such as signal quality, coverage
limitations, and the balance between cost and complexity. By addressing these considerations
strategically, retailers can maximize the effectiveness and value of VLC-based IPS in enhancing
operational efficiency and customer experience within the supermarket environment.

K. Wi-Fi-based

Wi-Fi, a widely adopted wireless technology, links devices to the internet via routers, offering a
range of 20 to 150 meters, extendable through overlapping Access Points (APs), making it ideal
for localization systems due to its global presence [214]. Its speed and range adhere to IEEE
protocol standards, with localization relying on ranging-based and fingerprinting-based methods.
WLAN, a local wireless network, operates in limited areas like schools and campuses, using
technologies such as Frequency Hopping Spread Spectrum (FHSS) and Direct Sequence Spread
Spectrum (DSSS). APs, acting as routers, connect to the internet, typically utilizing the 2.4 GHz
frequency with a range of 50-100 meters. Wi-Fi and WLAN-based IPS leverage these
technologies for precise and dynamic localization within enclosed spaces, revolutionizing indoor
navigation. By capitalizing on the proliferation of access points, routers, and beacons, these
systems enable the tracking and positioning of individuals and objects in real-time,
revolutionizing how we navigate and interact with indoor spaces. While Wi-Fi and WLAN are
frequently used interchangeably, in this context, we are addressing them under the Wi-Fi-based
category.
1) Wi-Fi Fingerprinting-Based

In fingerprinting-based methods, creating a fingerprint map and matching it to online fingerprint

data requires significant computing resources. However, indoor scenarios’ complexity often
leads to poor data quality during offline data collection for fingerprint databases. Therefore,
employing ML in fingerprint-based indoor positioning is preferable, reducing computing
resource usage without sacrificing accuracy. Additionally, this approach facilitates floor
distinction.

a) RSSI-Based

RSSI-based fingerprinting in Indoor Positioning Systems (IPS) serves to estimate the location of
mobile devices within indoor environments by measuring the power level of Wi-Fi signals
received from nearby Access Points (APs) or routers. While offering simplicity, low-cost
implementation, and compatibility with most Wi-Fi-enabled devices, RSSI-based methods are
susceptible to environmental changes like signal interference, multipath effects, and signal
strength variations due to obstacles. Nonetheless, RSSI-based fingerprinting remains widely
used, especially in environments where additional infrastructure installation isn’t feasible, such
as public spaces and shopping malls [97]. Recent years have seen an exploration of Deep Neural
Networks (DNNs) in RSSI-based IPSs, with systems utilizing DNNs to classify and regress Wi-
Fi user positions in indoor environments [99], [215], [216], [217]. For instance, A. Adege et al.
introduced a DNN-based indoor localization technique employing Wi-Fi [99], achieving high
accuracy by preprocessing RSS data with Linear Discriminant Analysis (LDA) and selecting the
strongest RSS values. Other techniques like Deep Positioning [215] combine RSSI and magnetic
field data for improved accuracy, while scalable approaches like the one proposed by K. S. Kim
et al. [216] utilize hierarchical DNN architectures for multi-floor classification with consistent
performance. Additionally, in [218], a Wi-Fi-based approach for localization of Mobile Nodes
(MNs) is introduced, employing RSSI to create location fingerprints and employing a
Convolutional Neural Network (CNN) on time series data to enhance accuracy, achieving a
mean error of 2.77m in predicting coordinates.

In [219], Roos et al. proposed an IPS. This grid-based Bayesian estimator achieved over 50%
accuracy within 1.5 meters in a small office. Similarly, [220] employed a Bayesian probabilistic
method for device localization. Battiti et al. in [23] developed a system using an MLP-based
classifier and OSS training, yielding less than 3 meters error with only 5 signal strength samples.
Its advantage lies in lower sensitivity to overfitting. [221] compares neural network and nearest
neighbor classifiers, achieving 72% accuracy within 1 meter. An MLP in [23] maps RSS data to
user location with an average accuracy of 2.3 meters, leveraging Wi-Fi infrastructure while
addressing privacy. [24] introduces the Modular Multi-Layer Perceptron (MMLP) for enhanced
accuracy in location estimation, managing uncertainty factors.

In [111], [112], the Horus system employs a joint clustering technique for probabilistic location
estimation. Each person’s location coordinate is treated as distinct classes, and Li is selected
based on the highest likelihood to minimize distance error. The experiments achieved over 90%
accuracy within 2.1 meters, with improved accuracy observed as the number of samples per
location increased. [90] presents an indoor positioning method using SVM and statistical
learning theory, showcasing SVM’s low error rate as a classifier and its regression version for
mobile user positioning. Another approach in [91] determines device location using RSSI from
APs, employing an SVM-based fingerprinting algorithm for real-time analysis. [25] introduces
the Discriminant-Adaptive Neural Network (DANN) for Wi-Fi client positioning, utilizing RSS
from APs to construct an accurate RSS-position relationship. Additionally, [222] proposes an
ANN-based approach for real-time target location detection and room type identification with
median positioning errors of 5.46m and 3.75m respectively. Finally, [223] presents a hybrid
WiFi and WSN-based approach using ANN, achieving an average distance error reduction to
1.05 meters, surpassing GA optimization.

In their work [224], Ladd et al. presented a grid-based Bayesian algorithm employing the 802.11
standard for robot localization. Using RSS data from 9 APs, the host employs a probabilistic
model to compute the position likelihood from a pool of locations, refining the results by
considering the mobile host’s limited maximum speed. Achieving over 83% accuracy within 1.5
meters, this method proves effective for robot localization. Similarly, [225] adopts a practical
Bayesian approach with the 802.11 architecture for topological localization within office
buildings, reducing training time without sacrificing accuracy. J. Zou et al. proposed a Wi-Fi
localization system [226] utilizing RSS with a deep regression model named DNN-CNN-DS,
comprising DNN, CNN, and Dempster-Shafer. An Auto-Encoder initializes the DNN weights,
optimized by minimizing the mean square error between model output and real location.
DFLAR, a device-free wireless localization and activity recognition technique [227], employs
deep learning to recognize activities and gestures based on the target’s influence on nearby
wireless links. Similarly, the approach in [228] enhances indoor Wi-Fi localization accuracy by
approximately 22% through improved contrastive learning and a parallel fusion network,
PaCNN-LSTM. Another system in [229] enhances WiFi indoor localization efficiency and
performance by utilizing PCA.

Another popular strategy in fingerprinting for indoor positioning involves utilizing fingerprint
images derived from RSSI data. Numerous approaches based on fingerprint image processing
and deep learning are presented in [214], [230], [231]. In [214], a dilated CNN is trained on RSS
images, and prediction errors are utilized to train an SVR model, verified using the
UJIIndoorLoc dataset [232]. Alternatively, in [230], Wi-Fi and magnetic signals are transformed
into fingerprint images, with a CNN learning mappings to actual positions, showcasing robust
learning and accuracy. Similarly, [231] introduces MFMCF, leveraging multi-pattern fingerprints
and various classifiers (KNN, SVM, RF) to enhance localization accuracy by constructing a
composite fingerprint set (CFS) with LDA from SSD, HLF, and RSS. Additionally, [233]
presents Wi-LO, an indoor localization system that boosts Wi-Fi-based accuracy by integrating
LTE

and magnetometer data, overcoming mismatches by combining different data types at each
location. Furthermore, another study [234] proposes an ML framework employing Bag-of-
Features and kNN classification, surpassing existing models in both simulations and real-time
experiments.
b) CSI-Based

CSI-based fingerprinting in IPS utilizes detailed information from Wi-Fi signals’ CSI to create
distinct fingerprints for various indoor locations. This approach offers high accuracy and
resilience in complex indoor environments but requires precise calibration and dense reference
point deployment [98]. Recent studies have extensively employed DL for CSI-based
fingerprinting in IPS [235], [236], [237], [238], [239], providing automatic feature learning,
noise resilience, adaptability to new environments, and real-time processing [240], [241], [242],
[243], [244]. Combining Wi-Fi signals with other methods can enhance accuracy, as seen in
[245], [246], where autoencoders reduced data dimensions, aiding in position estimation.
Additionally, systems like BiLoc [247] and [79] leverage DL to estimate location using 5-GHz
Wi-Fi CSI, achieving promising accuracy. Meanwhile, [248] introduces data rate (DR)
fingerprinting for passive localization, addressing challenges like low resolution and fluctuations
by employing various strategies such as transmission power levels and dynamic nearest
neighbors matching.

In [249], researchers explore utilizing Massive MIMO channels with CNN for localization,
leveraging the sparse structure of these channels to achieve fractional wavelength positional
accuracy. Another method, ConFi, employs CNN for Wi-Fi localization [250]. It treats time-
frequency metrics from CSI data as images and utilizes a 5-layer CNN for classification,
demonstrating robust performance with a 2.7 m localization error. Additionally, a CSI image-
based indoor localization method is introduced in [251], forming an RGB image with phase
differences and amplitudes from different antennas, and employing a CNN for classification.
Another CSI-based indoor fingerprinting system, DeepFi, trains all DNN weights as fingerprints
using deep learning in the offline phase, employing a greedy algorithm for complexity reduction.
In the online phase, a probabilistic data fusion method based on the RBF is utilized for location
estimation.

2) Wi-Fi Ranging-Based

A Wi-Fi IPS that utilizes trajectory CSI observed from predetermined routes instead of stationary
locations, addressing the limitations caused by multipath fading is proposed in [69]. The
proposed IPS employs a one-dimensional convolutional neural network-long short-term memory
(1DCNN-LSTM) architecture to leverage the spatial and temporal information of trajectory CSI.
Additionally, a generative adversarial network (GAN) helps enlarge the training dataset,
reducing the cost of trajectory CSI collection. All these studies are analyzed in Table. XI.

Wi-Fi signals usually encounter several challenges including signal propagation obstruction,
multipath effects, and non-line-of-sight scenarios, leading to inaccuracies. Achieving high
accuracy is difficult due to dynamic environments, infrastructure dependency, and privacy
concerns. Calibration and maintenance are crucial, and interference in crowded channels can
impact accuracy. Implementation complexity, cost, and power consumption pose further issues,
requiring advanced algorithms, hardware improvements, and robust signal processing techniques
for reliable Wi-Fi-based IPSs.

Wi-Fi technology, widely utilized for wireless connectivity, presents both opportunities and challenges when
considered for implementation in supermarkets, especially when viewed through the lens of ML-based tools.

**Pros:**

1. **Global Presence**: Wi-Fi's widespread adoption and global presence make it an attractive option for indoor
positioning systems (IPS) in supermarkets. Leveraging existing Wi-Fi infrastructure and routers allows for cost-
effective deployment and scalability.

2. **Versatility**: Wi-Fi IPS offers versatile localization methods, including fingerprinting-based and ranging-
based approaches, catering to diverse supermarket layouts and operational needs. ML-based algorithms enhance the
accuracy and efficiency of these localization methods, enabling precise tracking and positioning of individuals and
objects.

3. **Low-Cost Implementation**: Wi-Fi IPS, particularly RSSI-based fingerprinting methods, offers a low-cost
implementation compared to other indoor positioning technologies. ML algorithms optimize resource usage and
reduce computing requirements without compromising accuracy, making it an economically viable solution for
supermarkets.

4. **Real-Time Tracking**: By capitalizing on Wi-Fi signals from access points (APs) and routers, Wi-Fi IPS
enables real-time tracking and positioning of customers, staff, and inventory within the supermarket environment.
ML-based algorithms enhance the speed and responsiveness of these tracking systems, facilitating dynamic
navigation and personalized services.

**Cons:**

1. **Environmental Challenges**: Wi-Fi signals face challenges such as signal interference, multipath effects, and
signal propagation obstruction within the dynamic and crowded environment of a supermarket. These environmental
factors can introduce inaccuracies and variability in Wi-Fi-based positioning, requiring sophisticated ML-based
signal processing techniques to mitigate.

2. **Infrastructure Dependency**: Wi-Fi IPS relies heavily on existing infrastructure such as APs and routers for
signal transmission and reception. Any changes or disruptions to this infrastructure, such as network upgrades or
hardware malfunctions, can impact the performance and reliability of Wi-Fi-based positioning systems, requiring
ongoing calibration and maintenance efforts.

3. **Privacy Concerns**: Wi-Fi-based IPSs may raise privacy concerns among customers and staff, as they involve
tracking and collecting location data within the supermarket premises. Implementing robust privacy measures, such
as anonymization techniques and data encryption, is essential to address these concerns while complying with
privacy regulations.

4. **Complexity and Cost**: Implementing Wi-Fi-based IPSs in supermarkets requires advanced algorithms,
hardware improvements, and robust signal processing techniques to overcome challenges such as signal propagation
obstruction and non-line-of-sight scenarios. This complexity adds to the upfront costs and ongoing maintenance
expenses of the system, requiring investments in skilled personnel and technological infrastructure.
In conclusion, while Wi-Fi technology offers promising benefits for indoor positioning in supermarkets, including
global presence, versatility, and low-cost implementation, its implementation presents challenges related to
environmental factors, infrastructure dependency, privacy concerns, and complexity. By leveraging ML-based tools
and strategic planning, retailers can maximize the effectiveness of Wi-Fi-based IPSs in improving operational
efficiency and enhancing customer experience within the supermarket environment while addressing these
challenges.

Let's delve deeper into how Wi-Fi-based indoor positioning systems (IPSs) impact key performance indicators
(KPIs) such as accuracy, coverage, and cost & complexity within the context of implementing these systems in
supermarkets.

**Accuracy:**

- *Pros*: Wi-Fi IPSs offer high accuracy in indoor localization, especially with ML-based algorithms optimizing
localization methods like RSSI-based fingerprinting. These algorithms enhance the precision of Wi-Fi-based
positioning systems, allowing for accurate tracking and navigation of individuals and objects within the supermarket
environment.

- *Cons*: Despite advancements in ML algorithms, Wi-Fi-based IPSs may encounter challenges that affect
accuracy, including signal interference, multipath effects, and environmental factors like signal propagation
obstruction. These challenges require sophisticated signal processing techniques and continuous calibration to
maintain accuracy, adding complexity to system deployment and maintenance.

**Coverage:**

- *Pros*: Wi-Fi signals provide extensive coverage within supermarkets, utilizing routers and access points (APs)
strategically placed throughout the premises. This coverage ensures comprehensive monitoring and tracking of
individuals, staff, and inventory across various areas within the supermarket, facilitating efficient operations and
customer service.

- *Cons*: While Wi-Fi signals offer broad coverage, coverage consistency may be influenced by environmental
factors such as signal interference and obstacles within the supermarket environment. Ensuring consistent coverage
across all areas may require additional APs or routers, increasing deployment costs and complexity.
**Cost & Complexity:**

- *Pros*: Wi-Fi-based IPSs offer cost-effective solutions for indoor positioning in supermarkets, leveraging existing
Wi-Fi infrastructure and routers. This reduces upfront hardware costs and simplifies deployment, particularly when
utilizing fingerprinting-based methods that require minimal additional hardware.

- *Cons*: Implementing Wi-Fi-based IPSs in supermarkets requires careful consideration of infrastructure

dependency, privacy concerns, and ongoing maintenance costs. While leveraging existing Wi-Fi infrastructure
reduces hardware costs, optimizing signal processing algorithms and maintaining system accuracy may require
investments in skilled personnel and technological resources, increasing overall complexity and long-term costs.

In summary, Wi-Fi-based indoor positioning systems in supermarkets offer high accuracy, extensive coverage, and
cost-effective solutions, especially when combined with ML-based optimization techniques. However, challenges
related to signal interference, coverage consistency, and ongoing maintenance costs must be addressed to ensure the
successful implementation and operation of Wi-Fi-based IPSs in the dynamic environment of a supermarket.

I. Radio Frequency Identification (RFID)-based

An RF-compatible circuit facilitates electromagnetic trans- mission for data storage and retrieval. Essential
components of a basic RFID system include RFID readers, RFID tags, and elements for communication. The system
transmits and receives data within a predetermined radio frequency and protocol, operating in either passive or
active mode [188].

Passive RFID systems serve as alternatives to traditional bar code technology, being smaller, simpler, and more
cost- effective compared to active systems. These tags operate with- out a battery, reflecting the RF signal from the
reader to the receiver, with information added by modulating the reflected signal. However, their operational range
is limited to 1-2 meters, making them unsuitable for larger areas. Commonly

used frequency bands include LF, HF, UHF, and microwave frequencies [48]. Active RFID systems utilize similar
fre- quency ranges, featuring small transceivers that communicate their ID upon interrogation. These systems
typically offer a larger range and find applications like SpotON, a 3D location sensing system utilizing radio signal
strength analysis for object location detection [189].

In [192], a Particle Swarm Optimization Artificial Neural Network (PSO-ANN) algorithm for RFID indoor
positioning is proposed, using PSO to optimize ANN weights and thresh- olds. This cost-efficient RFID IPS is
compared with BPANN, ANN, and LANDMARC models. The approach establishes the relationship between RSSI
and tag position, enhancing accuracy. A Gaussian filter is employed for data processing, mitigating environmental
effects. Achieving an average posi- tioning error of 0.6482 m, the method leverages non-contact two-way
communication for data transfer. In a similar vein, [193] presents a Genetic Algorithm-Backpropagation Neural
Network approach, combining GA’s optimum searching with BPNN’s optimization for RSSI-based indoor
positioning. Also, [194] introduces a high-accuracy indoor location system using active RFID and neural network
classification. In [108], a Radial Basis Function Neural Network (RBFNN) with Virtual Reference Tags is employed
for RFID-based target localiza- tion, achieving a consistent accuracy with a positioning error of 0.472 m, while
selecting optimal network architectures.

In [195], an integrated wireless platform proposes an adaptive RFID-derived RSSI-based indoor location sensing
technique. Utilizing a fuzzy neural network architecture, the method adapts to environmental parameters. Active
readers and tags are used to enable long-distance transmission, achiev- ing less than 1 m positioning error with fewer
tags and readers. The approach in [196] introduces an intelligent tag strength prediction algorithm using
backpropagation learning in neural networks for RFID tag position detection. It predicts tag signal strength under
varying conditions and achieves over 90% ac- curacy in estimating target positions. While [197] presents an RFID
hybrid positioning method employing a neural network phased array antenna for indoor RF localization. Combining
AOA and RSSI, it scans the search plane using phased array antenna radiation beams, achieving a mean positioning
error of 0.32m for 10 locations.

In [198], an algorithm is presented for estimating target location in dynamic indoor environments like
warehouses with changing layouts. By utilizing RSSI and passive UHF tags as references, a trained RBFNN reduces
the Localized Generalization Error (L-GEM) for 2D warehouse positioning. Similarly, [199] employs multiple
neural networks and a genetic algorithm to estimate indoor positions from RSSI data collected from reference tags,
achieving an average error of 2.4 m and a maximum error of 5.21 m. Additionally, [200] explores active RFID for
positioning moving targets using Cluster-based Movable Tag Localization (CMTL) with kNN and ANN, achieving
an average positioning error of 0.77 m. These studies are compared in Table IX. Despite numerous approaches for
IPSs utilizing RFID technology, challenges persist, including fluctuations in signal strength, positioning
inaccuracies, environmental factors, and tag-related config- urations, underscoring the need to address these issues
for improved reliability and accuracy.

J. Bluetooth-based

Bluetooth, a wireless technology used for short-range com- munication among various devices, like computers and
smart- phones, employs affordable transceiver microchips and radio systems, ensuring minimal power consumption.
Its operational range, typically around 10-15 meters, varies based on factors such as materials and has a widespread
standard operating frequency of 2.4 GHz with a lower bit rate. Bluetooth-based indoor localization, favored for its
presence in mobile devices, cost-effectiveness, and energy efficiency, utilizes unique IDs for precise tag location [3],
proving valuable for tracking and room-specific tasks.

The Denoising Autoencoder-based Bluetooth Low Energy (BLE) indoor localization (DABIL) [201] employs a 3D
indoor positioning system utilizing Bluetooth data. Utilizing a denoising autoencoder, it extracts relevant
fingerprints from RSSI data to create a 3D reference point database. This method demonstrates improved vertical
and horizontal accuracies with a 1.27m positioning error. Neural networks are commonly utilized for Bluetooth-
based localization. For instance, in one approach [202], a cost-effective system utilizes neural net- works for user
orientation, achieving a 0.5m accuracy. Another

study [203] employs Bluetooth in phones for neighboring positions, using a 6-neuron hidden layer with a 17% error
rate. Recurrent networks in [101] focus on real-time Bluetooth localization with a 10m error. In [204], kNN
surpasses neural networks and SVM for indoor Bluetooth positioning with less than 1m error. Bluetooth RSS in
[205], employing CNNs, attains 93.33% accuracy and less than 1.4m error, suitable for multi-floor settings,
effectively covering large areas.

The study in [206] proposes a BLE-based indoor local- ization system utilizing Android devices to create contin-
uous radio maps with BLE ”iBeacons” and Wi-Fi access points. Stationary object localization is assessed using Wi-
Fi, BLE, and their combination, with optimal parameter selection through Weighted Nearest Neighbors (WNN).
Another study in [207] proposes a remote indoor positioning system using kNN analysis and a portable BLE tag,
applicable to tracking individuals in various scenarios. The demand for precise indoor localization arises from
superstores, smart homes, and disaster management needs. Addressing the complexity of indoor settings, [208]
describes an explainable indoor localiza- tion (EIL) technique employing BLE’s RSSI with a gradient boosting
machine. It achieves 98.04% accuracy within 1.5m in a superstore environment.

This study in [209] explores IoT-driven location-based services, particularly in indoor Bluetooth localization using
Bluetooth 5.1’s AOA function. They proposed a DL-based algorithm that fuses RSSI and AOA features through
PCA and KF. CNN extracts deep-level features from RSSI and AOA, followed by concatenation and Softmax layer
classification. The authors of [210] also proposed an indoor localization system using Bluetooth fingerprinting and
CNN. A unique approach transforms wireless signal data into images using a blurring technique to simulate signal
diffusion. Additionally, two-dimensional reduction algorithms, PCA and t-SNE, are compared. An evolutionary
algorithm configures the solution with varied transmission power levels. Results demonstrate a promising accuracy
of nearly 94%, highlighting the potential of this technique for enhanced indoor localization systems.

An improved RSSI-based fingerprinting method, employing data augmentation and ML algorithms for XY-position
identi- fication of user nodes relative to anchor nodes, is proposed in [211]. RF achieved a 96% test accuracy,
outperforming other techniques, ensuring precision, and compatibility with ML. In [212], a smartphone-based
indoor location method utilizing BLE Beacons’ RSSI values is suggested. ML is used to create a distance estimator
from RSSI readings. TensorFlow is employed to calculate intersection points of peripheral lines, facilitating position
estimation based on the geometric median. Additionally, [213] enhances accuracy by utilizing multiple an- chors
and radio channels. RF proves most effective, achieving over 99% classification accuracy. These methods are
compared in Table.X. Despite Bluetooth technology’s energy efficiency, device compatibility, and cost-
effectiveness in IPS, challenges like limited range, signal interference, accuracy variation, and privacy concerns
persist. Overcoming these challenges necessitates careful implementation and innovative solutions for optimal
performance in indoor positioning systems.
F. Magnetic Sensors-based

Magnetic sensors are integral to IPS, detecting Earth’s magnetic field changes indoors to ascertain device
position and orientation. Variations caused by objects or architecture create distinct patterns for
localization. Geomagnetic fingerprint- ing involves comparing real-time readings with a magnetic database.
Offering stability and simplicity, magnetic sensors are preferred in indoor settings with unreliable RF
signals. Machine learning, like recurrent neural networks, enhances accuracy by analyzing magnetic data
sequences, strengthen- ing IPS performance. Positioning and tracking systems using magnetic signals offer
high accuracy even in NLoS scenarios [162].

Existing magnetic positioning methods encounter chal- lenges in wide-area accuracy due to magnetic data
ambigu- ity, often requiring multiple sensors. Addressing this, [162] proposes an indoor system utilizing
distorted geomagnetic fields. Features from magnetic sequences are extracted, and fed into a trained neural
network alongside a magnetic map, yielding 2D locations with 80% accuracy. Similarly, [163] employs an
ANN with 5 hidden layer neurons for nonlinear input-output mapping, enhancing self-sensing active
magnetic bearing systems.

Existing RF-based indoor positioning algorithms are unsuit- able for large-scale areas like airports due to
proportional posi- tioning errors. A geomagnetic sensor-based method proposed in [164] uses stable
geomagnetic data for indoor localiza- tion. Object movement affects geomagnetic signals, and RNN models
track signal variations for target position detection. RNNs recognize time-varying sensor data sequences.
Training and testing occur using the indoor space’s magnetic field map, tuning hyperparameters.
TensorFlow and CUDA Toolkit are used, achieving positioning errors of 0.51m and 1.04m for medium and
large spaces, respectively. Another magnetic localization method is proposed using a multi-scale temporal
convolutional network (TCN) and LSTM in [165]. Time- series preprocessing enhances geomagnetic signal
discernibil- ity. TCN expands feature dimensions while preserving LSTM time-series characteristics. The
proposed stacking framework of multi-scale TCN and LSTM demonstrates effective indoor

magnetic localization. DeepML, another smartphone-based indoor localization system, employs LSTM
networks using magnetic and light sensors. Bi-modal images are generated via preprocessing and fed into
the network for training, enabling new device localization. It’s tested across varying environments,
showcasing its consistency [102].

In [166], the authors introduce an indoor localization test- bed, achieving a high accuracy of 98% through
XGBoost algorithms applied to magnetometer sensors and Wi-Fi access points, and examine classifier
performance with varying test data sizes. Studied in [167], [168], enhances smartphone indoor localization
by combining Wi-Fi RSSI and magnetic field data through DNN. It encompasses offline learning to extract
intrinsic features from multi-class fingerprints and an online serving phase. Instead of Wi-Fi signals, [169]
explored indoor localization using stable geomagnetic sensor signals. A DNN model and an RNN track
unique geomagnetic field signal sequences caused by object movement for positioning. Basic RNN and
LSTM versions are trained on magnetic field maps of medium and large-scale indoor testbeds. Sim- ilarly,
[170] utilizes smartphone sensors like magnetometer, accelerometer, and gyroscope for indoor localization.
It em- ploys fingerprinting based on magnetic flux intensity patterns, mitigating database updating and
device heterogeneity issues. ANN aids user state identification (walking/stationary) with 95% accuracy.
The study in [171] proposed a multi-sensor fusion approach to mitigate device dependency and enhance
accuracy. A DCNN recognizes indoor scenes to refine localiza- tion, while a magnetic field pattern
database minimizes device reliance. Modified K nearest neighbor (mKNN), pedestrian dead reckoning, and
an EKF further refine localization. A CNN-based IPS utilizing magnetic patterns (MP) for local- ization is
proposed in [172]. A database of MP is created, and CNN matches user-collected MP to estimate positions,
employing a voting mechanism for accuracy. An optimized geomagnetic positioning system using an
enhanced genetic algorithm (EGA) and extreme learning machine (ELM) is pro- posed in [173]. EGA
optimizes ELM’s parameters, achieving meter-level accuracy with robustness and faster construction. All
the papers mentioned in this section are analyzed in terms of complexity. scalability and cost in Table. VII.
However, magnetic sensor-based IPS encounters issues such as signal variability, device heterogeneity,
limited range, dynamic envi- ronment effects, and the need for accurate real-time updates, impacting the
system’s accuracy and reliability.
 BLE

Accuracy: Several studies demonstrate the potential for achieving high accuracy in
Bluetooth-based indoor localization. Techniques such as Denoising Autoencoder-based
Bluetooth Low Energy (BLE) indoor localization (DABIL) [201] showcase improved vertical
and horizontal accuracies with positioning errors as low as 1.27 meters. Neural networks
and k-nearest neighbors (kNN) algorithms have also been successful in achieving sub-
meter accuracy [202], [204]. Additionally, innovative approaches utilizing Bluetooth 5.1's
Angle of Arrival (AOA) function and deep learning algorithms show promising results,
with accuracies reaching nearly 94% [209], [210].

Coverage: Bluetooth technology offers widespread coverage within indoor

environments, making it suitable for large areas such as supermarkets. Studies
demonstrate effective coverage across multi-floor settings, enabling continuous radio
map creation and precise localization [206]. Furthermore, advancements in Bluetooth-
based techniques, such as the utilization of multiple anchors and radio channels,
contribute to enhanced coverage and accuracy [213].

Cost: Bluetooth technology is renowned for its cost-effectiveness, making it a favorable

choice for indoor positioning systems. The affordability of transceiver microchips and
radio systems ensures minimal power consumption, contributing to overall cost
efficiency [3]. Moreover, the implementation of ML-based algorithms utilizing Bluetooth
RSSI values and BLE Beacons offers a cost-effective solution for indoor localization [212].

Energy: Bluetooth Low Energy (BLE) technology ensures minimal power consumption,
making it energy-efficient for indoor positioning applications. This energy efficiency is
crucial for prolonging device battery life, especially in supermarket environments where
continuous operation is essential [3]. Techniques such as smartphone-based indoor
location methods utilizing BLE Beacons' RSSI values leverage energy-efficient protocols
to optimize energy consumption [212].

Computation: ML-based algorithms, particularly neural networks and kNN, are

commonly employed in Bluetooth-based indoor positioning systems for real-time
localization tasks [202], [204]. These algorithms facilitate efficient data analysis and
position estimation, ensuring swift updates of target locations across vast areas.
Additionally, advancements in ML-driven algorithms, such as explainable indoor
localization (EIL) techniques, contribute to improved computation efficiency and
accuracy [208].

Overall, Bluetooth-based techniques offer a promising avenue for enhancing indoor

positioning systems in supermarkets, addressing key considerations such as accuracy,
coverage, cost, energy efficiency, and computation. Despite challenges such as limited
range and signal interference, innovative solutions and careful implementation can
optimize Bluetooth-based IPSs for optimal performance in supermarket environments.

VLC

Accuracy: Visible light communication (VLC) technologies offer promising avenues for
achieving high accuracy in indoor positioning systems (IPSs). ML-based methods, such
as artificial neural networks (ANNs) and genetic algorithms, have shown significant
potential in enhancing accuracy. For instance, ANN-based VLC indoor positioning
systems have achieved mean position errors as low as 1.02 cm [95]. Additionally,
regression neural networks employed in 3D visible light positioning (3DVLP) have
demonstrated consistent accuracy with mean errors of 1.1 cm [100].

Coverage: VLC-based IPSs excel in coverage, leveraging the ubiquity of LED lighting
systems for indoor illumination. With the resilience to electromagnetic interference and
the ability to transmit location information through visible light, VLC technologies offer
extensive coverage within indoor environments, including supermarkets.

Cost: VLC-based IPSs are cost-effective, leveraging existing LED infrastructure for
communication. The affordability and simplicity of photo-diodes (PDs) make them
preferred receivers for VLC detection, contributing to overall cost efficiency [95].
Additionally, the utilization of ML algorithms in VLC-based localization systems
optimizes resource utilization and reduces operational costs.
Energy: VLC technologies, particularly when integrated with LED lighting systems, are
energy-efficient, ensuring minimal power consumption. This energy efficiency is crucial
for prolonged operation, especially in supermarket environments where continuous
operation is essential. Techniques such as ANN-based VLC indoor positioning systems
optimize energy consumption while maintaining high accuracy [95].

Computation: ML-driven algorithms, including ANN and deep neural networks (DNN),
play a vital role in processing data and optimizing computation efficiency in VLC-based
IPSs. These algorithms enable real-time data analysis and position estimation, facilitating
swift updates of target locations across vast areas. Additionally, advancements in ML
algorithms, such as Position Estimation Deep Neural Networks (PE-DNN), contribute to
enhanced computation efficiency and accuracy [114].

Despite the promising attributes of VLC-based IPSs, challenges such as signal blockage,
lighting conditions, and line-of-sight constraints persist. Addressing these challenges
through innovative algorithms, optimized hardware, and integration with lighting
systems is essential for improving accuracy and feasibility in supermarket environments.
A comprehensive comparison of VLC-based IPSs, considering factors like complexity,
scalability, and cost, is necessary for guiding future implementations in supermarket
settings.

WIFI

Accuracy: Wi-Fi-based indoor positioning systems (IPSs) offer high accuracy through
fingerprinting and ranging-based methods, aided by machine learning (ML) algorithms.
Deep neural networks (DNNs), convolutional neural networks (CNNs), and recurrent
neural networks (RNNs) have demonstrated significant improvements in accuracy,
achieving mean errors as low as 1.02 cm [95] and 2.77 m [218] in predicting coordinates.
Bayesian estimators, support vector machines (SVMs), and discriminant-adaptive neural
networks (DANNs) have also showcased high accuracy within indoor environments
[111], [25], [222].

Coverage: Wi-Fi-based IPSs leverage the extensive coverage of Wi-Fi and wireless local
area networks (WLANs) within indoor environments, ensuring comprehensive coverage
even in large areas like supermarkets. By utilizing access points (APs) and routers, these
systems enable dynamic localization and tracking of individuals and objects in real-time,
revolutionizing indoor navigation.
Cost: Wi-Fi-based IPSs offer cost-effective solutions by leveraging existing Wi-Fi
infrastructure for communication. The affordability and ubiquity of Wi-Fi-enabled
devices contribute to overall cost efficiency. ML algorithms optimize resource utilization
and reduce operational costs, making Wi-Fi-based IPSs economically viable for
implementation in various indoor environments, including supermarkets.

Energy: Wi-Fi technologies, known for their energy efficiency, ensure minimal power
consumption in IPSs. By utilizing Wi-Fi signals for localization, these systems maintain
high accuracy while conserving energy, essential for prolonged operation in
supermarket environments where continuous operation is crucial.

Computation: ML-driven algorithms, including DNNs, CNNs, and RNNs, optimize

computation efficiency in Wi-Fi-based IPSs, enabling real-time data analysis and
position estimation. Bayesian estimators, SVMs, and adaptive neural networks further
enhance computation efficiency while maintaining high accuracy within indoor
environments.

Despite the promising attributes of Wi-Fi-based IPSs, challenges such as signal

propagation obstruction, multipath effects, and non-line-of-sight scenarios persist.
Calibration, maintenance, and interference in crowded channels are critical
considerations for achieving reliable and accurate Wi-Fi-based IPSs in supermarket
environments. Advanced algorithms, hardware improvements, and robust signal
processing techniques are essential for overcoming these challenges and realizing the
full potential of Wi-Fi-based IPSs in supermarkets.

MAG

Accuracy: Magnetic sensors, pivotal in indoor positioning systems (IPS), detect Earth’s
magnetic field changes, offering stability and simplicity in unreliable RF signal
environments. Machine learning, particularly recurrent neural networks (RNNs),
enhances accuracy by analyzing magnetic data sequences, ensuring high precision even
in non-line-of-sight (NLoS) scenarios [162]. Proposals such as distorted geomagnetic
field utilization [162] and RNN-based tracking of signal variations [164] demonstrate
significant improvements, achieving positioning errors as low as 0.51m [164].

Coverage: Magnetic sensor-based IPSs provide comprehensive coverage, particularly in

large-scale areas like supermarkets where RF-based algorithms may encounter
challenges. By tracking unique geomagnetic field signal sequences caused by object
movement, these systems ensure accurate positioning even in dynamic environments
[169], [170]. Proposals incorporating multi-sensor fusion [171] and CNN-based
localization [172] further enhance coverage by mitigating device dependency and
refining localization accuracy.

Cost: Despite challenges such as signal variability and device heterogeneity, magnetic
sensor-based IPSs offer cost-effective solutions for indoor localization. Optimized
algorithms utilizing enhanced genetic algorithms and extreme learning machines
achieve meter-level accuracy with robustness and faster construction [173], ensuring
cost-efficient implementation in various indoor environments, including supermarkets.

Energy: Magnetic sensor-based IPSs exhibit energy efficiency, crucial for prolonged
operation in dynamic indoor environments like supermarkets. Smartphone-based
systems employing LSTM networks and magnetic sensors optimize energy consumption
while maintaining consistency across varying environments [102], ensuring reliable real-
time updates without compromising system performance.

Computation: Machine learning algorithms, particularly RNNs and deep neural networks
(DNNs), optimize computation efficiency in magnetic sensor-based IPSs. Techniques
such as smartphone sensor fusion [170] enhance computation efficiency while
preserving high accuracy in indoor localization. Additionally, proposals integrating CNNs
for magnetic pattern recognition [172] demonstrate effective computation optimization
for accurate position estimation.

While magnetic sensor-based IPSs offer promising solutions for indoor localization,
challenges such as signal variability and limited range persist, impacting system
accuracy and reliability. Addressing these challenges through innovative algorithms and
robust real-time updates is essential for advancing magnetic sensor-based IPSs and
realizing their full potential in supermarket environments.