IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 7, NO.

4, OCTOBER-DECEMBER 2020 2271

Mobi-IoST: Mobility-Aware Cloud-Fog-Edge-IoT

Collaborative Framework for Time-Critical
Applications
Shreya Ghosh , Student Member, IEEE, Anwesha Mukherjee , Student Member, IEEE,
Soumya K. Ghosh , Senior Member, IEEE, and Rajkumar Buyya , Fellow, IEEE

Abstract—The design of mobility-aware framework for edge/fog generated by IoT devices, cloud computing plays a significant
computing for IoT systems with back-end cloud is gaining role. However, the cloud-only set-up is not an energy-efficient
research interest. In this paper, a mobility-driven cloud-fog-edge
collaborative real-time framework, Mobi-IoST, has been and delay-aware solution for handling such a high volume of
proposed, which has IoT, Edge, Fog and Cloud layers and exploits data. To address this problem, edge and fog computing have
the mobility dynamics of the moving agent. The IoT and edge been introduced [2]. On the other hand, seamless connectivity
devices are considered to be the moving agents in a 2-D space, due to the mobility of IoT devices is a crucial factor to process
typically over the road-network. The framework analyses the the data in the remote cloud servers. For time-critical applica-
spatio-temporal mobility data (GPS logs) along with the other
contextual information and employs machine learning algorithm tions such as health care, connection interruption and conse-
to predict the location of the moving agents (IoT and Edge devices) quently the increase in delay in delivering the processed
in real-time. The accumulated spatio-temporal traces from information, result in poor Quality of Service (QoS). If the
the moving agents are modelled using probabilistic graphical device gets disconnected due to mobility, the delivery of the
model. The major features of the proposed framework are: processed data/ information becomes a challenge. This neces-
(i) hierarchical processing of the information using IoT-Edge-Fog-
Cloud architecture to provide better QoS in real-time applications, sitates a hierarchical infrastructure, where each layer (IoT,
(ii) uses mobility information for predicting next location of the edge, fog or cloud) either accumulates, stores and processes
agents to deliver processed information, and (iii) efficiently handles the information for reducing the delay. On the other side,
delay and power consumption. The performance evaluations yield movement traces, i.e., time-stamped location information of
that the proposed Mobi-IoST framework has approximately moving agents (say, mobile-users or client) are accumulated
93% accuracy and reduced the delay and power by approximately
23–26% and 37–41% respectively than the existing mobility-aware on a large scale from GPS-enabled smart phones or IoT devi-
task delegation system. ces. This spatio-temporal movement information opens up
Index Terms—Cloud computing, Edge computing, Fog diverse opportunities to explore the intent of movement [3],
computing, Internet of Things (IoT), Mobility analytics, Spatio- [4] and thus fostering varied location based services, namely,
temporal data. efficient package delivery [5], traffic resource management
etc. Internet of Spatial Things (IoST) brings IoT in the spatial
context [6]. As discussed before, mobility or continuous
I. INTRODUCTION change of locations of users is a challenging issue in task dele-
gation or data offloading. However, analysis of these mobility
T HE advancements of Internet of Things (IoT) have mani-
fested significant improvements on the quality of human
lives in varied aspects [1]. To facilitate real-time applications,
information helps to explore the intent of the move and subse-
quently extracts the frequent movement path of a user in dif-
ferent contexts. If the probable location sequences of an agent
high-end processing and storage units are required. For com-
in the near future can be predicted from the historical mobility
putation and storage of these large volume of raw data
information, then an effective and delay-aware solution for a
time-critical application can be provided.
Manuscript received April 2, 2019; revised August 9, 2019; accepted Sep-
tember 4, 2019. Date of publication September 16, 2019; date of current ver- To address the above-mentioned challenges, we propose a
sion December 30, 2020. This work was supported in part by the Department of Cloud-Fog-Edge based collaborative framework for the proc-
Science and Technology, Government of India through sponsored research proj- essing of IoT data and delivering the result based on mobility
ect to Indian Institute of Technology Kharagpur, India and in part by Melbourne-
Chindia Cloud Computing (MC3) Research Network. Recommended for accep- analysis to reduce the delay. We have considered a hierarchi-
tance by Dr. A. Puliafito. (Corresponding author: Shreya Ghosh.) cal mobility-based infrastructure composed of four layers: IoT
S. Ghosh, A. Mukherjee, and S. K. Ghosh are with the Department of Computer layer, edge layer, fog layer, and cloud layer. Nowadays smart
Science and Engineering, Indian Institute of Technology Kharagpur, West Bengal
721302, India (e-mail: shreya.cst@gmail.com; anweshamukherjee2011@gmail. phone has become a popular medium for ubiquitous Internet
com; skg@cse.iitkgp.ac.in). access and varied user-specific IoT applications are accessible
R. Buyya is with the CLOUDS Laboratory, School of Computing and through smart phones. These mobile devices serve as edge
Information Systems, The University of Melbourne, VIC 3010, Australia
(e-mail: rbuyya@unimelb.edu.au).
devices and may frequently change the locations. The edge
Digital Object Identifier 10.1109/TNSE.2019.2941754 layer contains such edge devices i.e. mobile devices. The fog
2327-4697 ß 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 2272 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 7, NO. 4, OCTOBER-DECEMBER 2020

a fog device. The respective RSUs can actuate the signal syn-
chronizing mechanism such that am can reach avoiding the
congested route as well as without waiting in the traffic sig-
nals. Although the scenario is motivated by the dysfunctional
public health system and limited access to improved transpor-
tation and medical care in the rural areas, specifically in devel-
oping countries, such as India, Mobi-IoST is beneficial for any
time-critical applications. For instance, in the time of emer-
gency, a police-vehicle needs to commute with minimal delay
avoiding the congested regions of a city. Mobi-IoST predicts
the less congested route by analyzing the traffic states in real-
time and notifies the RSUs of the route. These RSUs actuate
the signal synchronizing mechanism such that the vehicle can
Fig. 1. Mobi-IoST for health care application (a): Ambulance sends health reach the destination avoiding the congested route as well as
data from IoT devices to RSU. (b): RSU sends result with the location without waiting in the traffic signals. The hierarchical place-
information to cloud. (c): Cloud predicts the nearby health care centre, shortest ment of IoT, edge, fog devices and cloud servers in Mobi-
path and helps to actuate traffic signal.
IoST framework facilitates an effective and delay-aware solu-
tion for several time-critical applications. We believe that
layer contains the fog devices such as RSUs (Road Side Unit) Mobi-IoST will act as a foundation of mobility aware network
which are large cell base stations. While the IoT and the edge resource management for varied location-based service plan-
devices change their locations, the RSU and the cloud data ning in real-time.
centers of the framework have static locations. The raw data
generated in the IoT layer is sent to the edge layer, which is B. Contributions
connected with the fog layer. The fog layer is connected with The focus of our work is to develop a cloud-fog-edge col-
the cloud layer where high-end processing and mobility analy- laborative framework which facilitates real-time IoT informa-
sis tasks are performed. tion processing and delivery of results based on the mobility
information analytics. The key contributions can be summa-
A. Motivating Scenario
rized as follows:
We have considered a well-known time-critical application
Mobi-IoST (Mobility-aware Internet of Spatial Things)
in the domain of health care, where the proposed Mobi-IoST is designed for information processing and delivering
framework can be deployed. The pictorial representation of result based on the prediction of user’s current location.
this use-case is shown in Fig. 1. The framework exploits the mobility knowledge of the
Suppose a patient, travelling in a vehicle (am), needs con- agents to predict the probable user location and delivery
tinuous monitoring of her/his vital health parameters such as of processed information at low delay and low power
blood-pressure, pulse-rate, body-temperature etc. which are consumption of the user-device.
collected using IoT devices and the raw data are sent to the
A novel mobility modelling network has been proposed
RSU through a client application. The RSU processes the to explore the movement patterns of the user. The huge
information based on functional model pre-defined by medical amount of spatio-temporal trajectory data is stored effi-
experts and sends the current status as normal/abnormal to the ciently along with other contextual information in the
client-app. If any abnormality is detected, the RSU sends cloud data centre.
the data to the cloud to find out the nearest health centre. In
A real-time mobility prediction module has been
the developing countries like India there is a scarcity of super- designed to predict the location sequences of the user
speciality hospitals at rural regions, and the ambulances are effectively.
also not equipped with good medical facilities and there is a
The experimental results demonstrate that the proposed
rare possibility of presence of a medical expert inside the system has outperformed other existing approaches in
ambulance. In such circumstances, the proposed framework accuracy and takes much less time to learn the patterns.
can provide a preliminary support for continuous health moni- The simulation results also demonstrate that the pro-
toring, as well as can suggest nearby health centres in case of posed framework reduces the delay in delivering infor-
adverse situation. Given the current location and health-data mation and power consumption of the mobile device
feed from the RSU, the cloud can suggest the nearby hospital. (user-device) compared to the existing mobility-aware
On the other side, based on the route followed by the vehicle, task delegation approach.
the probable health centre also gets notified. Further, the To the best of our knowledge, this work is the first attempt
mobility analysis module of cloud can help to reduce the com- to utilize the movement knowledge to enhance the QoS for
muting time of the vehicle by predicting less congested path facilitating time-critical IoT applications. The rest of the paper
in the road-network. This can be achieved when cloud analy- is structured as follows. Section II briefs the existing work in
ses the traffic states (congestion, traffic breakdown etc.) of the related areas. The proposed framework, Mobi-IoST, is dis-
roads and notifies the RSUs of the path. The RSU will work as cussed in Section III. The delay and power consumption

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 GHOSH et al.: MOBI-IOST: MOBILITY-AWARE FRAMEWORK FOR TIME-CRITICAL APPLICATIONS 2273

models are discussed in Section IV. Section V presents the serialization of session information has been discussed. Fur-
experimental and simulation results. The paper is concluded ther, there is a need of a mechanism where the user mobility
in Section VI along with future directions. will be predicted and result will be delivered at the optimum
delay and power consumption of the mobile device.
Given the abundance of mobility/trajectory traces (GPS log)
II. RELATED WORK
of individuals, there are several research initiatives to extract
The IoT refers to the connection of embedded devices knowledge or meaningful information from the huge amount
within an existing Internet infrastructure where the devices of trajectory traces. Several works are reported to model and
are uniquely identified and the computing environment is cre- predict next location from movement traces such as GPS log,
ated [1]. The raw data collected by IoT devices are processed check-in data or social network information [17]–[19]. There
inside the cloud servers. However, storing and processing of are challenging applications, namely, urban land-use classifi-
the raw data inside the remote cloud enhances the delay and cation from taxi-traces [20], daily activity-sequence recom-
energy consumption. To overcome this, fog computing has mendation [21] categorizing users in an academic campus
been introduced [2]. The raw data of IoT devices are proc- [22]. It is well known that human movement traces follow spa-
essed inside the fog device instead of the remote cloud to tio-temporal regularity. In this regard, Song et al. [23] provide
reduce the delay and energy consumption. However, during a high degree of spatio-temporal uniformity by mining move-
data processing connection interruption becomes a challenge ment traces of 50,000 people for a period of three months. All
if the client is a mobile device. IoT has several sub-domains of these studies depict that since people follow some spatio-
depending on its applications e.g. Internet of Multimedia temporal regularity in their movement history, an appropriate
Things (IoMT), Internet of Health Things (IoHT), Internet of and effective mobility pattern modelling can help to facilitate
Vehicles (IoV) etc [6]. IoST is a new sub-domain of IoT, several location-aware services.
which focuses on spatial data management [6]. In fog-based To this end, the Mobi-IoST framework aims to deliver
IoT, the switch, routers etc. work as fog devices for faster mobility-driven efficient data processing in cloud-fog-edge
processing of the raw data collected using IoT devices. The based IoT setup to facilitate intelligent decision making in
mobile device that usually works as an edge device, is a real-time. To the best of our knowledge, none of the existing
connector between the IoT devices and the network. However, works have clearly depicted the significance of mobility-aware
the resource hindrance is a major difficulty for these handheld service provisioning framework in fog based IoT. In Mobi-
devices. Therefore, the cloud servers have to be used by IoST, the movement pattern modelling and location prediction
mobile devices to store their data [7]. The mobile devices also approaches are novel propositions which deliver result in real-
offload heavy computations to the cloud servers. Energy and time. Moreover, the experimental observations and perfor-
latency in offloading have been focused on several existing mance analysis show the effectiveness of Mobi-IoST in terms
approaches [8], [9]. Fog computing has also provided solu- of accuracy, delay and power consumption. In summary,
tions for reducing delay and energy in the processing of designing and deploying an end-to-end mobility driven frame-
IoT data [2]. In fog computing, a hierarchical architecture is work for efficient data processing in IoT setup is a challenging
followed, where the intermediate devices between the end issue in the present era.
node (or fog devices) and cloud servers participate in data
processing [2]. The edge devices allow users to connect with
III. MOBI-IOST FRAMEWORK
the network and transfer data accordingly to a network which
is external to the user. For transcoding massive amount of The hierarchical structure of Mobi-IoST is represented in
video at scale, a cloud and edge computing based collabora- Fig. 2. Fig. 3 depicts the overall flow of the framework, Mobi-
tive system has been proposed in [10]. For balancing the traffic IoST. IoST or Internet of Spatial Things deals with IoT data
and computing load, a method has been discussed in [11], along with spatial perspective. As depicted in Fig. 2, in the
where the IoT devices are allocated to the base station or fog bottom layer several IoT sensors such as accelerometer, GPS,
nodes to reduce the latency. temperature, blood-pressure, proximity sensors capture appli-
To address the task offloading issue in vehicular network, cation specific data. These IoT sensors are either present
edge computing has been used in [12]. The mobile edge com- within the edge devices or connected with the edge devices,
puting servers are deployed inside the road side access points namely, mobile phone, vehicles, which change their locations.
for offloading tasks. However, the use of access points may When any of these edge devices needs assistance, it contacts
not be energy-efficient if exhaustive computations have to be the current RSU (the RSU under which it currently belongs).
performed and there are a large number of users. Based on In this work, RSU is used as fog device and it is capable of
user mobility, an opportunistic computation offloading small scale processing. If the processing is beyond the compu-
method has been discussed in [13]. Based on information tational capability of the RSU, then it delegates the task to the
gain, task allocation in spatial crowdsourcing has been dis- cloud. The top layer of the hierarchical structure consists of
cussed in [14]. A fog based architecture of spatial crowdsourc- cloud servers, which store spatial data, specifically, mobility
ing has been proposed in [15], where privacy-aware task traces, location-specific information, city-structure (POI
allocation and data aggregation have been focused. In [16], placements and other contextual information). The cloud proc-
task offloading to cloud and delivery of result based on essing unit executes the task and sends the result to the RSU,

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 2274 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 7, NO. 4, OCTOBER-DECEMBER 2020

Fig. 2. Hierarchical placement of IoT, Edge, Fog devices and Cloud in Mobi-IoST framework.

where the tasks are application-specific such as controlling the A. Exploring Movement Semantics From Trajectory Traces
signaling mechanism or notifying the nearest health-care cen- This section presents the methodology to model movement
ter. Here, the resources can be efficiently managed by this patterns and predicts the next location sequences efficiently
framework: movement analysis module can predict variation and timely manner. Location prediction of moving agents,
of travel demand apriori and notify the RSUs accordingly, such as, people, vehicles is a challenging task for varied loca-
while the RSUs can decide about the dissemination of resour- tion-based services [20]. Specifically, in our work, location
ces (traffic or network) efficiently. prediction helps to locate the moving agent’s locations in near
The major modules of the framework are: (i) movement pat- future and subsequently data is sent to the appropriate RSU.
tern modelling, (ii) predicting next location sequences, (iii) Whenever a mobile device gets connected to a RSU, the GPS
delivery of result after processing in a timely manner. Finally, log of the mobile device is extracted and stored in the cloud
the experimental and simulation results yield the effectiveness dynamically.
of the framework.

Fig. 3. Working model of the proposed framework, Mobi-IoST.

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 GHOSH et al.: MOBI-IOST: MOBILITY-AWARE FRAMEWORK FOR TIME-CRITICAL APPLICATIONS 2275

It may be noted that location prediction depends on several iterative reverse geo-coding technique to extract nearest
factors, namely, day of the week, time-slot of a day and road- landmark of the stay-point.
structure. The first step of movement behaviour modelling is After the addition of semantic information with the raw
to find out the frequent pattern followed by the users in varied traces, a trajectory trace takes the form:
contexts. For example, the path followed by an individual dif- < pa ; ta ; Residential > ; TrajW ½ðpi ; ti ; ex Þ; ðpiþ1 ; ti þ dt; ex Þ;
fers significantly in weekends compared to his/her weekdays’ ðpiþ2 ; ti þ 2  dt; ex Þ; ; < pb ; tb ; SuperMarket > ; TrajW ½ðpj ;
trajectory signature. Moreover human movements follow tj ; ey Þ; ðpjþ1 ; tj þ dt; ey Þ; ðpjþ2 ; tj þ 2  dt; ey Þ; ; < pc ; tc ;
some intent [3] and extracting the purpose behind any move is Residential > :
the fundamental step to predict next location effectively. Few Here, pa ; pb and pc are three stay-points with geo-tagged
preliminary concepts which are used in this paper are defined information residential building and supermarket. TrajW
as follows: stores the route information followed by the trajectory, where

GPS log (G): GPS log is the collection of time-stamped ex ; ey are the road-segments of the road-network of the study
latitude, longitude information. The GPS trajectory or region.
trace is formed by connecting the location information Processing of Large Mobility Datasets in Cloud: With
on increasing time-ordering. the advances in sensor technologies and the proliferation of
Trajðp1 ; . . . ; pn Þ : < p1 ðlat1 ; lon1 Þ; t1 > ! < p2 ðlat2 ; smartphones, a huge amount of mobility traces are generated
lon2 Þ; t2 >    ! < pn ðlatn ; lonn Þ; tn > , where t1 < by moving agents. One of the major challenges is to analyze
t2 <    < tn . the vast amount of data due to computational complexity and

Stay-Point (S): Stay-point of a trajectory is defined as a storage limitations. To this end, we propose to migrate the
location (typically, polygon), where the moving agent computation of mobility analysis and storage of movement
stops for a time-value dt and dt > tthresh , and all the traces in the cloud for faster response. It may be noted that
GPS points within dt reside in the area areastay of the the locations and coverage areas of the RSUs need to be
polygon, where areastay < arear . Here, polygon is a maintained in the cloud storage such that it can predict the
data-type which represents spatial data [24] and tthresh , next location of the moving agent to determine the appropri-
arear are time-threshold and coverage-area threshold ate RSU, which will serve the agent at that time. Here, large
for detecting stay-points from the trajectory respec- cell base stations [26] are the RSUs. The macrocell base sta-
tively. In our analysis, we have considered the parame- tion is referred to as macro RSU and microcell base station
ter values as, tthresh ¼ 12mins and arear ¼ 2 km2 . is referred to as micro RSU. The coverage area of macro

POI and Geo-tagged Trajectory: Point-of-interest (POI) RSU and micro RSU are 1–20 km and 200 m–1 km respec-
of a GPS location denotes the nearby landmark of a tively [26]. The framework is implemented in Google Cloud
location, such as, residential area, supermarket etc. We Platform (GCP) by utilizing several storage and computa-
have followed the POItaxonomy 1 to extract such POI tional components of GCP. In our framework, cloud storage
information using Google Place API. Geotagged trajec- is of four types:
tory is generated by appending the geo-tagged informa- – Grid based storage of the study region: The study
tion of the stay-points within the trajectory. region is segmented into uniform hexagonal grids and

Trajectory window (TrajW ): Trajectory window stores information, such as, road structure or POIs, RSUs are
the location sequence information between two such associated with each such grids. Our proposition is to
stay-points in an uniform sampling rate. segment the spatial region into grids such that each grid
Augmenting Semantic Information with GPS log: Human encloses the coverage area of at least one micro RSU.
movement semantics can be analysed if additional information The grid-segmentation process initiates with the loca-
such as, POI, road-network structure and stay-point informa- tion of one RSU. Suppose, the location of the RSU (say,
tion are appended with the raw GPS traces. RSUi ) is p ¼ ðx; yÞ and the length of the side of the
– Road network of the study region is extracted from hexagon (gi ) is a ¼ 8 m. An iterative process is deployed
OpenStreetMap (OSM)2. The road network is repre- until the complete study region is segmented with hexag-
sented by a directed graph R ¼ ðV; EÞ, where e  jEj onal grids. In the first iteration, center points of the 6
denotes the road-segments of the region and v  jV j is neighbouring grids of gi are calculated and subsequently,
the intersection points of such road segments. Map- the neighbouring hexagonal grids are constructed.
matching algorithm [25] has been deployed, which con- In the next step, Geohash code of all hexagonal grids
siders both geometric and topological structure of the are computed. Geohash code of the grids represent the
road-network to associate the road-segments along with spatial location on the earth surface using unique alpha-
the trajectory traces. numeric strings. Cloud Spanner of GCP is used to store
– Each stay-point of the trajectory is geo-tagged with the these information which supports horizontal scalability.
nearby POI location. Here, we have implemented the – Road network information storage: This module stores
the road network information, namely, connections
among different road-segments and road-type (highway,
1
https://developer.foursquare.com/docs/resources/categories/ lane etc.). The information is stored in an adjacency
2
OpenStreetMap: https://www.openstreetmap.org matrix format, where each vertex maintains a list of

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 2276 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 7, NO. 4, OCTOBER-DECEMBER 2020

outgoing edges (outdegree of the vertex). The data-type both spatial location and temporal span of a visit-sequence to
of the list is polyline, which is a spatial-data type [24] model FPN of user movement graph. Each node (stay-points:
and represents the road-segments on the map. v  V ) of the network is conditionally independent of its non-
– RSU information storage: It stores the list of RSUs descendants given its parent node (PaðvÞ). Suppose, a visit-
along with the unique-id, coverage area, location (lati- sequence is given as V ¼ ðV1 ; . . . ; VN Þ, the probability distri-
tude, longitude) and other information such as, type bution is computed as follows:
(micro RSU or macro RSU) etc. Cloud BigQuery of Y
N
GCP is utilized to store the road network information P ðV Þ ¼ P ðVi jPaðVi ÞÞ (1)
and RSU information as well. i¼1
– Frequent path storage: The frequent path followed by
individual moving agents are extracted and modelled in FPN captures the dynamic nature of the mobility information
our work. Details are presented in Section III-A1. Cloud by representing multiple copies of the spatial-information, one
Bigtable of GCP is utilized to store the road network for each time-slice Vt ¼ ðV1;t . . . ; Vd;t Þ. Subsequently, the
information. transition distribution from one state to other (P ðVtþ1 jVt Þ) is
The computational cost of the mobility traces is huge since computed from two time-slice Bayesian network. The spatial
it deals with time-series data with very high sampling rate. location information (Vt ) is typically divided into two sets,
The key challenge is to reduce the processing time of the loca- namely, unobserved state variables (St ) and the observed state
tion prediction, and therefore, an efficient scheme is required. variables ( Lt , in our case, location information from RSUs).
Here, we have deployed a hash-based indexing scheme, where The joint probability distribution is calculated by unrolling
nearby locations are stored in the subsequent buckets of the two time-slice Bayesian networks:
hash-table.
P ðS0 ; . . . ; ST ; L0 ; . . . ; LT Þ ¼
1) Movement Behaviour Modelling: In this section, we
discuss how movement behaviour of users can be modelled to Y
T (2)
explore the frequent paths followed by them in different con- P ðS0 ÞP ðL0 jS0 Þ P ðSt jSt1 ÞP ðLt jSt Þ
t¼1
texts. The process of semantic enrichment of GPS log of users
has already been discussed in Section III-A. Here, we propose It may be noted that we have represented the stay-points as
User movement graph, a multi-layer graphical model to model grid-location and transition from one state to another state sig-
the users’ movement patterns from the spatio-temporal con- nifies that the agent is moving from one grid to another.
text. The objective to use the multi-layer network is that Next, we deploy a spatio-temporal trajectory clustering
human movement patterns typically depend on temporal var- (TrajCS) on FPN which captures the signature or frequently
iations (weekdays or weekends, morning or evening), road visited paths of the individual. The process follows a hierar-
networks and stay-points. All of these information need to be chical top-down approach, and based on the distance measure
encoded and interconnections of the information cannot be new clusters are generated and appended in the list. The trajec-
properly captured in a single layer. tory clustering distance measure is computed as follows:
User movement graph (MG): User movement graph is TrajCSðSi ; Sj Þ ¼
defined as MG ¼ ðN; L; laÞ, where N denotes the nodes, L
8
denotes the links and label is represented by la. The user > 0 ifði ¼¼ 0Þ
>
>
movement graph has four labels: >
> orðj ¼¼ 0Þ
>
>

Road network: The layer 1 consists of road network >
< TrajCSðSi1 ; Sj1 Þ ifððSi ¼¼ Sj Þ
information, where nodes are intersection points of (3)
>
> þC  MinðTScorei ; TScorej Þ andðsiþ1 Þ 6¼ ðsjþ1 ÞÞ
road-segments. >
>
>
> MAXðTrajCSðSi1 ; Sj Þ;

RSU network: The RSU information (location and cov- >
>
:
erage area) is stored in layer 2. TrajCSðSi ; Sj1 ÞÞ ifðsi 6¼ sj Þ

Stay-point information: The stay-point information
including location and type of stay-point are stored in where Si represents a set of locations, si denotes one GPS
layer 3. point of the set Si and C is the parameter to augment the prob-

Frequent path: The movement paths frequently fol- ability of the path taken. Tscore computes the temporal similar-
lowed by the user is stored in layer 4. ity between two different stay-points of the trajectory and
It may be noted that each layer is interconnected with each TrajCS method is recursively called for extracting the signa-
other. As the construction of layer 1, layer 2 and layer 3 are ture pattern. The proposed distance measure appends temporal
straight forward, we discuss the frequent pattern mining pro- information with the conventional Longest Common Sub-
cess of layer 4 in detail. Sequence (LCSS) clustering method [27]. Algorithm 1
The frequent path network of layer 4 is represented by prob- describes the basic steps of generating FPN from the trajec-
abilistic graphical model or Dynamic Bayesian network tory traces of agents.
FPNðV; E; Þ where V is the set of stay-points, E denotes the This section describes how users’ frequent movement pat-
direction of visit among different stay-points and  is the net- terns are extracted and stored along with other contextual
work quantify parameter. In this work, we have considered information. Furthermore, the trajectory clustering and multi-

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 GHOSH et al.: MOBI-IOST: MOBILITY-AWARE FRAMEWORK FOR TIME-CRITICAL APPLICATIONS 2277

layer graphical model help to effectively model the movement update process updates the result based on the current input
behaviour of agents in the cloud server. from the mobile device. It may be noted that the order (k)
of the markov chain is dependent on the user’s frequent
Algorithm 1: Frequent path mining - A trajectory clustering movement pattern and extracted from FPN of user move-
approach ment graph (MG).
Input: Set of trajectory T , stay-points S
Output: Frequent path network < FPNðV; E; Þ > Algorithm 2: Location prediction - map and update process
" Frequent path network for each individual Input: User movement graph MG, Present location s, Trajectory log T
1: clus; S; V; E NULL; Output: <s0 ; Edge  list> "sequence of next location sequences
2: for each trajectoty window tr 2 T do 1: function MAPPER(s; MG; T ) :
3: for each unvisited stay-point s 2 S do 2: j geo  hascodeðsÞ " extract the grids where trajectory
4: V:appendðsÞ " Create new node points placed
5: visited s 3: E 0 extract patternðMG; jÞ " extract the frequent
6:  : CPT Create ConditionalProbabilityTableðsÞ trajectory patterns within the grid
7: t extractTemporalðsÞ " Temporal information 4: for all ti 2 T do
of the stay-point 5: L predictLoc[xðti ; sÞ] " HMM based prediction
8: E:appendðgenEdgeðS; tÞÞ " genEdge creates directed 6: p ComputeProbðs; arraylist½L; tÞ " Compute
edge based on frequency of the visit and temporal information transition probability for all patterns in partcular time-stamp
9: end for 7: dist ComputeTrajCSðE 0 ; ti Þ " Compute the
10: end for trajCS distance for each such pattern
11: for each path trp 2 FPN do " Path represents the 8: s0 SORT ðarraylist½p; arraylist½distÞ " Predicted
successive sequence of stay-points location with maximum probability
12: Ne ExtractTrajW ðtrp ; FPNÞ " Extract 9: end for
trajectory-windows within spatio-temporal locality 10: Print < s0 : arraylistðE00 Þ >
13: D computeLCSSðNe; trp Þ 11: function Update(s; arraylist½s0 ) : " If sudden change
14: if D > thresh then in mobility pattern observed
15: Ignore Ne 12: for all ti do
16: end if 13: for all ea 2 arraylist½s0  do
17: if D  thresh then 14: Append trajectory window Tre containing ea in FPN
18: Create new cluster clust 15: ModifyCPT ðea ; ti Þ " Modify CPT of all nodes
19: clustt :appendðNe; trp Þ " Appending new cluster with outgoing/incoming edge ea
in the list 16: ModifyProbðea ; ti Þ " Modify transition matrix
20: visited Ne of all sequences containing ea
21: Modify CPT ðclustt Þ " Modify the CPT of all 17: Mapperðs; MG; Tre Þ
nodes in clustt calculating the frequency of visit 18: end for
22: end if 19: end for
23: end for

P ðsi jsi1 ; si2 ; . . . ; s1 Þ ¼ P ðsi jsi1 ; . . . ; sik Þ (4)

2) Next Location Sequence Prediction From Mobility
Information: This helps to calculate the probable path visit
by the user apriori. The location prediction task is formulated In the next step, forward algorithm [28] is deployed to find
as follows: Given the historical observations of a moving out the sequences of stay-points, given as:
agent m and the current location L at time t, predict the "
agent’s anticipated location sequences (S) following d time- X
seq max Y k

instances P ðL k
jxÞ ¼ P ðLðjÞjSi ðjÞÞ
i¼1 j¼1
Typically, the task is formulated as information retrieval # (5)
task considering the fact that people’s movement patterns fol-
P ðSi ðjÞjSi ðj  1Þ; Si ðj  2Þ; . . . ; 1Þ
low spatio-temporal regularity and effective movement behav-
iour modelling leads to accurate location prediction. Here, we
have deployed Hidden Markov model (HMM) (say x) based where, seqmax and Si ðjÞ represent the maximum number of
prediction technique with two kinds of stochastic variables, hidden state sequences and hidden states. Here, model is
state variables (hidden) and observable variables. Each indi- represented as k-order markov chain where the next loca-
vidual’s movement is modelled as kth order Markov chain and tion depends on k recent observations. Next, a variant of
the transition from one place to another place is modelled verterbi algorithm using time-relationships is deployed to
based on MG and x. discover the possible sequences of states. The transition
Algorithm 2 describes the basic steps of location predic- and emission probabilities of x are computed by adjusting
tion. The first step map locates the current location (grid the model parameters. An iterative version of forward
location) of the moving agent and then the model predicts backward algorithm is implemented to produce the sequen-
the sequences of locations based on the context and finally, ces effectively.

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 2278 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 7, NO. 4, OCTOBER-DECEMBER 2020

Three types of location prediction tasks have been carried Algorithm 3: Working Model of Mobi-IoST
out in this work: (i) location sequence prediction in a specific Input: Raw data received from IoT devices
time-threshold, (ii) predicting appropriate POI (say, health- Output: Result after processing the raw data
care center) and the path and finally, (iii) given the destination 1: mobile device receives raw data from the IoT devices
and present location of the agent predicting the path with less 2: if mobile device is able to process the data then
commuting time based on the traffic states of the road-net- 3: mobile device works as edge device and processes the data
work. The location sequence prediction in specific time- 4: else
5: mobile device forwards the data to the fog device RSU under
thresholds is computed directly from x on MG. The POI and
which it is currently present
path prediction is carried out by overlapping the road-network
6: if current load of the RSU < maximum load capacity
structure (layer 1 of MG) and frequent path pattern (layer 4 of of the RSU then
MG). Finally, given source and destination, the markov model 7: go to step 11
is used along with the traffic-state of the region, where s1 8: else
and sn are specified. It may be noted that our algorithm 9: go to step 34
(Algorithm 2) is self-adaptive, i.e., the update function (also, 10: end if
refer ’Update’ arrow of Fig. 3) changes the modelling algo- 11: if RSU is able to process the data then
rithm in case any of the prediction result fails. 12: RSU processes the data
13: if the mobile device is still connected then
B. Delivery of Result After Data Processing 14: RSU delivers the result to the device
15: else
The IoT devices are connected with the edge device, e.g. the
16: RSU forwards result to the cloud along with the device
sensors within the mobile device. With the increasing availabil-
ID and the request ID
ity of smartphones, we have considered mobile devices as the 17: cloud predicts current location of the mobile device
edge devices. If the mobile device is able to process the raw from the mobility information using Algorithm 1 and 2
data received from the IoT devices, it does the same by working 18: cloud identifies the RSU serving the predicted location
as an edge device and generates the result. Otherwise, the 19: cloud forwards the result to the predicted RSU along
mobile device sends the data to the RSU, which will act as a with the device ID and the request ID
fog device. Each RSU maintains a look-up table, which holds 20: if the mobile device is connected with the predicted
the mobile device IDs present under its coverage. The Interna- RSU then
tional Mobile Equipment Identity (IMEI) number is considered 21: RSU sends the result to the mobile device
as the mobile device ID. The RSU after receiving the raw data 22: else
23: RSU sends a feedback to the cloud that the mobile
from the mobile device, checks its current load and ability to
device is not present in its coverage
process the data. In this regard, two cases appear as follows:
24: cloud after receiving the feedback stores the result

If the RSU’s current load is same as the maximum load it 25: if the mobile device gets connected with a RSU
can handle or an exhaustive computation is required to then
perform which is beyond the capability of the RSU, it 26: mobile device requests for the result to the RSU
forwards the data to the cloud along with the device ID with the request ID
and the request ID. After processing the data, the cloud 27: RSU forwards the request to the cloud
finds the current location of the device based on the 28: cloud sends the result to the RSU and updates
user’s geo-location information (refer Section III-A). the mobility information
Based on the current location of the device, the cloud 29: RSU sends the result to the mobile device
identifies the RSU under which the mobile device is cur- 30: end if
31: end if
rently located. The cloud sends the result to the RSU
32: end if
along with the device ID and the request ID. The RSU
33: else
forwards the result to the mobile device. 34: RSU sends the raw data along with the device ID and

If the RSU is able to process the raw data and its current the request ID to the cloud
load is less than the maximum load it can handle, the 35: cloud processes the data
RSU processes the data and sends the result to the mobile 36: go to step 17
device. However, as the device is in mobility, it may 37: end if
be possible that the RSU finishes processing and the 38: end if
mobile device moves away. In such a case, the RSU
sends the result to the cloud along with the request ID In our approach as the mobility information is updated
and the device ID. The current location of the device is dynamically, the probability of the presence of the mobile
predicted by the cloud based on the user’s geo-location device under the predicted RSU is high. However, if the
information. Based on the current location of the device, device losses connection with the network for a long duration,
the cloud identifies the RSU under which the mobile there is a probability that the mobile is not located under the
device is currently located. The cloud sends the result to coverage of the predicted RSU. In such cases, after receiving
the RSU along with the device ID and the request ID. the result from the cloud, the predicted RSU sends feedback to
The RSU forwards the result to the mobile device. the cloud that the mobile device is not present under its

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 GHOSH et al.: MOBI-IOST: MOBILITY-AWARE FRAMEWORK FOR TIME-CRITICAL APPLICATIONS 2279

TABLE I The total data transmission delay between mobile device and
SYMBOLS USED IN POWER AND DELAY CALCULATION
RSU is given as:
Tt ¼ Tmr þ Trm (8)
The data processing delay inside the RSU is given as:

Tpr ¼ Dc =dpr (9)

The total delay for data transmission and processing is:

Ttot ¼ Tt þ Tpr (10)

The mobile device while located at point i transmits the data

to the RSU and requests for processing. Let the radius of the
RSU’s coverage area is R and the last visited point of the
mobile device inside the RSU is k. The delay in movement
from point i to point k is given as:

X
k1
Tik ¼ ððDiðiþ1Þ Þ=ui Þ (11)
i¼1

where ui is the velocity of the mobile device at location point i

coverage. By this time, if the mobile device gets connected and Diðiþ1Þ is the distance between two consecutive location
with a RSU, it will send a request for the result with the points i and i þ 1. If Tik > Ttot , the mobile device is still
request ID to the RSU. The RSU then forwards it to the cloud. inside the coverage of the RSU. Hence, the RSU will deliver
The cloud then sends the result to the mobile device through the result to the mobile device. Hence, the round-trip delay is
the RSU and updates the user mobility information accord- given as:
ingly. Algorithm 3 summarizes the steps of the working model Tdel1 ¼ Ttot (12)
of the proposed framework Mobi-IoST. As we observe in the
proposed system the cloud after analysing the user mobility The power consumption of the mobile device during this
information, finds out the probable RSU currently serving the period is given as:
device and accordingly delivers the result. Hence, intelligent
Pdel1 ¼ Tt Pa þ Tpr Pi (13)
decision making is performed by the cloud, which makes the
system efficient. As the RSU performs data processing instead of the cloud, the
transmission delay is reduced. Hence, the delay in delivering
IV. DELAY AND POWER CONSUMPTION IN MOBI-IOST the result is reduced in our system. Accordingly, the power con-
The mobile device transmits the data to the RSU with which sumption of the mobile device is also reduced. Else if
it is currently connected and requests for providing the result Tik  Ttot , the mobile device has moved to the coverage of
after processing. According to our strategy, either RSU or another RSU. Hence, the previous RSU will forward the result
cloud performs data processing based on the complexity of to the cloud along with the request ID and the mobile device
the computation required for processing the data and the cur- ID. The cloud contains the mobility information of the mobile
rent load of the RSU. Here, two cases are considered as dis- devices. Using the location prediction strategy described in the
cussed previously: Section III-A2, the cloud will find out the current probable

Information processing inside the RSU location of the mobile device. Let t is the time instant when the

Information processing inside the cloud mobile device is at location i. The cloud finds out the location
The symbols used in the delay and power calculation of the point visited by the mobile device at time instant ðt þ Ttot Þ, and
proposed model, are defined in Table I. the RSU which is currently serving the device. The cloud deliv-
ers the result to the selected RSU, which forwards the result to
A. Delay Model for Information Processing in RSU the mobile device. In this case, the round-trip delay is:

The uplink data transmission delay between mobile device Tdel21 ¼ Ttot þ ðDr =Uptrc Þð1 þ frc Þ
and RSU is given as: (14)
þ Tm þ ðDr =Dwtrc Þð1 þ fcr Þ
Tmr ¼ ð1 þ fmr Þ ðDc =Uptmr Þ (6)
where Tm is the delay for determining the current location of
The downlink data transmission delay between mobile device the device and correspondingly the RSU currently serving the
and RSU is given as: device, based on the mobility information of the user. The
power consumption of the mobile device during this period is
Trm ¼ ð1 þ frm Þ ðDr =Dwtmr Þ (7) given as:

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 2280 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 7, NO. 4, OCTOBER-DECEMBER 2020

Pdel21 ¼ Tt Pa þ ðTpr þ ðDr =Uptrc Þð1 þ frc Þ Ttn ¼ Tt þ Trc þ Tcr (22)
(15)
þ Tm þ ðDr =Dwtrc Þð1 þ fcr ÞÞ Pi
The data processing delay inside the cloud is given as:
If the device is not connected with the selected RSU, the RSU
Tc ¼ Dc =dpc (23)
will send a feedback to the cloud. If the mobile device sends
request to a RSU for the result, then the cloud will deliver the The cloud after processing the data, predicts the current loca-
result to the device through the current RSU. In this case, the tion of the user by analysing the mobility information and
round-trip delay is given as: accordingly the RSU currently serving the mobile device. After
that the cloud sends the result to that RSU along with the device
Tdel22 ¼ Tdel21 þ Tf þ Tr1 þ Tr2 þ ðDr =Dwtrc Þð1 þ fcr Þ ID and the request ID. Hence, the round-trip delay is:
(16)
Tdel31 ¼ Ttn þ Tc þ Tm (24)
where Tf is the delay for sending feedback from the RSU to
where Tm is the delay in predicting the current location and
the cloud, Tr1 is the delay for sending request by a mobile
the RSU serving the device currently. The power consumption
device for result to the RSU, under which the device is pres-
of the mobile device during this period is given as:
ent, and Tr2 is the delay for forwarding the request by the
RSU to the cloud. The power consumption of the mobile Pdel31 ¼ Tt Pa þ ðTrc þ Tcr þ Tc þ Tm Þ Pi (25)
device during this period is given as:
If the device is not connected with the selected RSU, the RSU
Pdel22 ¼ Pdel21 þ Tr1 Pa þ ðTf þ Tr2 will send a feedback to the cloud. If the mobile device sends
(17)
þ ðDr =Dwtrc Þð1 þ fcr ÞÞ Pi request to a RSU for the result, then the cloud will deliver the
result to the device through the current RSU. In this case, the
If ps and pu are the probabilities of the presence and non-pres- round-trip delay is given as:
ence of the mobile device under the predicted RSU respec-
tively, the round-trip delay is given as: Tdel32 ¼ Tdel31 þ Tf þ Tr1 þ Tr2 þ ðDr =Dwtrc Þð1 þ fcr Þ (26)

Tdel2 ¼ ps Tdel21 þ pu Tdel22 (18) where Tf is the delay for sending feedback from the RSU to
the cloud, Tr1 is the delay for sending request by a mobile
The power consumption of the mobile device during this device for result to the RSU, under which the device is pres-
period is given as: ent, and Tr2 is the delay for forwarding the request by the
RSU to the cloud. The power consumption of the mobile
Pdel2 ¼ ps Pdel21 þ pu Pdel22 (19) device during this period is given as:

However, though we have considered the case that the mobile Pdel32 ¼ Pdel31 þ Tr1 Pa þ ðTf þ Tr2
device may not be present under the coverage of the predicted (27)
þ ðDr =Dwtrc Þð1 þ fcr ÞÞ Pi
RSU, the probability of this case is very low, because the
cloud is dynamically maintaining the user mobility informa- If ps and pu are the probabilities of the presence and non-pres-
tion. As the user current location and the current RSU serving ence of the mobile device under the predicted RSU respec-
the mobile device is predicted in our system, the delay in tively, the round-trip delay is given as:
delivering the result is reduced. Accordingly, the power con-
sumption of the mobile device is reduced. Tdel3 ¼ ps Tdel31 þ pu Tdel32 (28)

B. Delay Model for Information Processing in Cloud The power consumption of the mobile device during this
period is given as:
From the previous subsection the total delay in data trans-
mission between RSU and mobile device (Tt ) has been deter- Pdel3 ¼ ps Pdel31 þ pu Pdel32 (29)
mined using equation (10). Now, if cloud performs data
processing, then the uplink data transmission delay between However, as the cloud is dynamically maintaining the user
RSU and cloud is given as: mobility information, the probability of the case that the
mobile device is not present under the coverage of the pre-
Trc ¼ ð1 þ frc Þ ðDc =Uptrc Þ (20) dicted RSU is very low. As in the proposed model the RSU
under which the user is currently present is predicted, the
The downlink data transmission delay between RSU and cloud delay in delivering the result is faster and correspondingly the
is given as: power consumption of the mobile device is reduced.
Tcr ¼ ð1 þ fcr Þ ðDr =Dwtrc Þ (21) V. PERFORMANCE EVALUATION
Therefore, the total data transmission delay between mobile The performance analysis has been carried out in following
device and cloud is given as: aspects: (i) movement pattern modelling, (ii) next location

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 GHOSH et al.: MOBI-IOST: MOBILITY-AWARE FRAMEWORK FOR TIME-CRITICAL APPLICATIONS 2281

TABLE II
PERFORMANCE COMPARISON OF MOVEMENT MODELLING MODULE OF MOBI-IOST WITH BASELINE METHODS

(and sequence) prediction and (iii) route prediction given the Markov chain [30], Convolutional Neural Network (CNN)
source and destination pair. Approach [19] and Spatio-temporal Recurrent Neural network
(ST-RNN) [18]. It may be noted that the trajectory modelling
A. Mobility Dataset modules of all of the cited works have been implemented with
our dataset for better comparison and to depict the effective-
The mobility dataset3 is collected from 100 mobile users
ness of our framework.
from their GPS-enabled smart phones and Google Map time-
One of the major challenge of the proposed framework is to
line for 6 months in the Kharagpur and Kolkata region of
reduce the delay in delivering the processed information, and
India. The dataset consists of timeseries data of GPS traces
therefore if new GPS trace comes, the system should be able to
with the total time-duration of 26,8041 hours. The GPS points
re-learn the pattern effectively. Table II shows the performance
are logged in a high-sampling rate of 60-75 secs.
measurements (accuracy, learning and re-learning time) com-
pared to six baseline methods. The cardinality (number of tra-
B. Experimental Setup
jectory-windows × day) of the test data of each agent for
We aim to demonstrate the efficacy of Mobi-IoST with learning and re-learning are 18  150 and 6  20 respectively.
the real-life mobility dataset. Typically, accuracy, recall Vlachos et al. [27] propose non-metric similarity function
and F-measure are used to evaluate the performance of LCSS by computing the similarity between trajectory segments.
Mobi-IoST. Six baseline methods are implemented to com- However, it only calculates the topological or geometrical simi-
pare with the proposed Mobi-IoST approach. 70% of the larity ignoring the semantics of the trajectories. Semantic
movement traces are used for modelling, 20% and 10% for enrichment of trajectories and modelling to predict next loca-
testing and validating respectively. The location sequence tion has been studied in [17]. Mingqi et al. utilizes the Bayesian
prediction task is evaluated in different time-scales, from network place classifier [29] to categorize the semantic of stay-
5 mins to 60 mins. The path prediction task has been points. Cheng et al. [30] model the check-in sequences of indi-
carried out in seven different time-bins (commuting time) viduals using markov chain for personalized POI recommenda-
(i) 10 mins, (ii) > 10 and 15 mins, (iii) > 15 and 20 tions. Recently researchers are devoted to deploy neural
mins, (iv) > 20 and 30 mins, (v) > 30 and 40 mins, networks [18], [19] to predict the next location sequences accu-
(vi) > 40 and 45 mins and (vii) > 45 and  50 mins. The rately. Karatzoglou et al. [19] has utilized CNN for modelling
accuracy measure is represented by the path similarity semantic trajectories. However, it is observed from our experi-
between the road-segments in the predicted path and the mentation that this method does not perform well for the infre-
actual query trajectory. For this purpose, the trips are quent or less-occurring locations in the trajectories and when
divided into seven classes based on their commuting time. the sequence of the stay-points are larger than 8. Since [18] is
For each class, the tenfold cross validation policy has been capable to model individuals’ mobility pattern in different con-
deployed where all trips within the same class are randomly texts by considering semantic information and spatio-temporal
divided into ten folds, where nine folds are utilized for periodic behaviour as well, we have carried out a detailed com-
training and one fold for validation. It guarantees that any parison study with [18]. It is observed from Table II that our
trip in the validation set will not appear in the training set. framework, Mobi-IoST outperforms all other baselines, except
Next, the prediction accuracy for all the trips in the valida- ST-RNN by approximately 10–18%. Mobi-IoST not only pro-
tion set are computed and the average value of the accuracy vides next location prediction based on some prediction tech-
measure for all seven classes are reported. nique (such as, CNN, RNN or Markov-model), rather it models
individual’s movement patterns over days, captures the fre-
C. Movement Analysis quent path followed in several contexts and makes the next
The performance measurement of the movement analysis location sequence prediction. Further, it is observed that the
module have been carried out by three measurements, namely, learning and re-learning rates of CNN and ST-RNN are signifi-
accuracy, recall and F-measure. Apart from that, we evaluate cantly higher by 10–20 mins and 14–24 mins than Mobi-IoST
the performance of the movement behaviour modelling frame- respectively. These measurements are important for our case,
work by comparing with six baseline methods, semantic tra- since the system needs to incorporate any sudden movement
jectory modelling [17], Bayesian network [29], LCSS [27], pattern change of user effectively. The neural network based
methods are costly in terms of re-learning and stability. In sum-
3 mary, the deep learning architectures used in the existing works
Sample dataset available: https://drive.google.com/drive/folders/1BpM-
K3clH6XYpSHkFe12aGsG8n1AclI4?usp=sharing are computationally intensive, and it is shown that such deep

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 2282 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 7, NO. 4, OCTOBER-DECEMBER 2020

Fig. 4. Recall values for location prediction based on time-stamp value Fig. 6. Accuracy percentage for prediction trajectory sequences.
(min).

Fig. 5. F-measure values for location prediction based on time-stamp value Fig. 7. Accuracy values for path prediction given source and destination.
(min).

time-stamp values respectively. The results indicate that

architecture may not be beneficial for time-critical applications, Mobi-IoST not only provides accurate predictions, but per-
where a delay-aware solution is necessary. forms better than ST-RNN [18] while the time-stamp values
Figs. 4–6 show the performance metrics of Mobi-IoST with increase as shown in Figs. 4 and 5. The reason behind the con-
the existing work [18]. The accuracy to predict stay-point sistent prediction result with increased time-stamp value is
information is represented in Fig. 6. It may be observed that that we have considered both spatial location and temporal
Mobi-IoST has accuracy of 88% to 95% for trajectory stop span of a visit-sequence to model the proposed FPN of user
sequences ranging from 2 to 10. There is a significant drop movement graph. The prediction algorithm is capable to pre-
of accuracy percentage from 93% to 80% of [18] in the same dict next location sequences based on both recent k locations
set-up. The key reason behind this observation is the proposed and the time-duration of each stay-points. Thus, the frame-
movement modelling named User movement graph, where work is suitable to predict longer sequences of locations more
user movement pattern is modelled in a multi-layer graphical effectively than existing work. Fig. 7 represents the accuracy
model. The frequent path network (FPN) (layer 4 of the User of predicting path given source and destination, where x-axis
movement graph) learns the movement paths frequently fol- is the commuting time. The experimental set-up is discussed
lowed by the user deploying Dynamic Bayesian network. in Section V-B. All of the performance measure values are
Thus, the model captures all sequences of paths visited in computed based on the correct number of grid-id predictions.
different spatial and temporal contexts. Next, in the proposed The accuracy percentage lies in the range of 89% to 95:3% for
model, k-order markov chain is used to effectively model the seven time-bins.
spatio-temporal regularity of the trajectory sequences, where The performance evaluation is carried out to validate
the next location depends on k recent observations. On the whether Mobi-IoST is suitable to deliver the processed informa-
other side, the existing work [18] use deep architecture to cap- tion to the user-device based on user mobility prediction. To
ture the spatial and temporal contextual information, however assist users and make intelligent decisions in time-critical appli-
fall short to predict long sequences of trajectory stay-point or cations, it is very crucial to model and predict users’ next loca-
stop points. Fig. 4 and Fig. 5 show the recall and F-measure tions apriori based on varied spatio-temporal contexts. It has
values of location sequence prediction in different time-scale, been observed that our framework has outperformed in all of
from 5 mins to 60 mins. Since the major aim of this work the performance measurements compared to the baseline meth-
is the efficient delivery of service while the agent is on ods. The experimental results present that to predict user’s next
move, the experimental evaluation based on the time-stamp location Mobi-IoST takes approximately 4.23–9.02 seconds,
value is justified. We have found the recall and F-measure val- which is used in Section V-D to determine the delay and power
ues in the range of 0.95 to 0.81 and 0.96 to 0.78 for twelve consumption in Mobi-IoST.

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 GHOSH et al.: MOBI-IOST: MOBILITY-AWARE FRAMEWORK FOR TIME-CRITICAL APPLICATIONS 2283

TABLE III
SIMULATION PARAMETERS

Fig. 9. Power consumption of mobile device in proposed and existing

methods (first case).

(s) and watt (W) respectively. The delay and power consump-
tion of the mobile device in case of Mobi-IoST are compared
with the existing mobility-aware task delegation method [16].
This is observed that for the considered parameter values the
delay in Mobi-IoST and existing method [16] are approximately
0.02–0.03s and 0.03-0.06s respectively (see Fig. 8). This is also
observed that for the considered parameter values the power
consumption of mobile device in Mobi-IoST and existing
method [16] are approximately 2–3.5 mW and 2.5–4.5 mW
respectively (see Fig. 9). In our approach the RSU works as a
resourceful fog device and performs the data processing. If the
device moves to the coverage of another RSU, the cloud pre-
Fig. 8. Round-trip delay in proposed and existing methods (first case). dicts the current RSU based on user mobility information. In
conventional method, the cloud performs data processing and
D. Delay and Power Consumption the user receives the result through the RSU. However, if the
The mobile device sends data to the RSU for processing. The user gets disconnected due to movement to another RSU, the
RSU/cloud performs processing and sends back the result to user has to access the cloud to retrieve the result by serializing
the mobile device. The mobile device may move to the cover- session information [16]. But in our approach, the cloud itself
age of another RSU before acquiring the result. In such cases, sends the result to the RSU, that is currently serving the device.
the connection interruption period is considered 10–30 sec. The RSU then forwards the result to the mobile device. Hence,
MATLAB2015 is used for the simulation. The parameter val- the delay and power consumption of the mobile device in the
ues considered in this analysis are presented in Table III. proposed system Mobi-IoST are less than the existing system
In this analysis we consider the following two cases: [16]. This is observed that Mobi-IoST reduces the delay and

Information processing inside the RSU power by approximately 23–26% and 37–41% respectively

Information processing inside the cloud than the existing method [16].
In the first case, we have considered health parameter data In the second case, we have considered video data is trans-
transmitted by the mobile device. Blood pressure level (systolic mitted by the mobile device. The RSU after processing the
and diastolic), body temperature, pulse rate and ECG data are video data, sends back the processed data to the mobile device.
considered. The RSU works as fog device and has functional The amount of transmission to serve each user request is con-
model (pre-defined by the medical experts) to perform the proc- sidered 2–20 MB. The round-trip delay and power consumption
essing. Based on the input (health parameter values, health pro- of mobile device in the proposed approach, are presented in
file of the user/patient and ambience parameter values) the RSU Fig. 10 and Fig. 11, and compared with the existing mobility-
executes the functional model and predicts the health status. based task delegation method [16]. This is observed that for the
The RSU after processing the health data, sends back the current considered parameter values the delay in Mobi-IoST and exist-
health status (normal/abnormal) as result to the mobile device. ing method [16] are approximately 5–20s and 10–50s respec-
If abnormality is detected, then the parameters which seem to tively (see Fig. 10). This is also observed that for the considered
be abnormal are also notified in the result. The amount of data parameter values the power consumption of mobile device
transmission to serve each user request is considered 70–90 KB. in Mobi-IoST and existing method [16] are approximately
The round-trip delay and power consumption of the mobile  1:5 W and 0.5–3.5 W respectively (see Fig. 11). In our
device (user-device) in the proposed approach, are presented in approach the cloud performs the data processing. After that
Fig. 8 and Fig. 9. The delay and power are measured in second based on user geo-location information, the cloud predicts the

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 2284 IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, VOL. 7, NO. 4, OCTOBER-DECEMBER 2020

result to the client mobile device becomes a challenge if the

client frequently changes location. In this paper, we have
proposed a real-time cloud-fog-edge IoT collaborative frame-
work, namely Mobi-IoST, for efficiently delivering the proc-
essed information to the user-device based on user mobility
prediction and intelligent decision making. The mobile device
acts as an edge device, and the RSU is used as fog device for
processing the raw data collected by the mobile device from
the IoT devices. If the user changes location and gets discon-
nected, the RSU forwards the result to the cloud. The cloud
analyses the mobility pattern and delivers the result accord-
ingly. The mobility prediction module primarily stores the
movement traces and models the frequent path followed by
the individual in different contexts and deploys a hidden mar-
Fig. 10. Round-trip delay in proposed and existing methods (second case). kov model based location predictor for efficiently predicting
the location sequences. The real-life data of movement traces
yield approximately 10–18% improvement compared to other
existing methods. Moreover, the simulation results demon-
strate that the proposed fog computing framework reduces the
delay and power by approximately 23–26% and 37–41%
respectively than the existing mobility-aware task delegation
system. The Mobi-IoST framework is quite generic and can be
extended to capture the cellular data usage patterns from such
time-series data and prediction of location sequences may
help to appropriately manage the power and bandwidth related
resources.

REFERENCES
[1] J. Gubbi, R. Buyya, S. Marusic, and M. Palaniswami, “Internet of
Things (IoT): A vision, architectural elements, and future directions,”
Future Gener. Comput. Syst., vol. 29, no. 7, pp. 1645–1660, 2013.
[2] M. Chiang and T. Zhang, “Fog and IoT: An overview of research
Fig. 11. Power consumption of mobile device in proposed and existing opportunities,” IEEE Internet Things J., vol. 3, no. 6, pp. 854–864,
methods (second case). Dec. 2016.
[3] S. Ghosh and S. K. Ghosh, “Thump: Semantic analysis on trajectory
traces to explore human movement pattern,” in Proc. 25th Int. Conf.
current location of the user and the RSU currently serving the Companion World Wide Web, 2016, pp. 35–36.
device. The cloud forwards the result to the RSU, which then [4] S. Ghosh and S. K. Ghosh, “Exploring human movement behaviour
sends back the result to the mobile device. However, in the based on mobility association rule mining of trajectory traces,” in Proc.
Int. Conf. Intell. Syst. Des. Appl., Springer, 2017, pp. 451–463.
existing method, the cloud performs data processing and [5] C. Chen et al., “Crowddeliver: Planning city-wide package delivery
the user receives the result through the RSU after accessing the paths leveraging the crowd of taxis,” IEEE Trans. Intell. Transp. Syst.,
cloud. Moreover, if the user gets disconnected and moves to the vol. 18, no. 6, pp. 1478–1496, Jun. 2017.
[6] K. A. Eldrandaly, M. Abdel-Basset, and L. A. Shawky, “Internet of spa-
coverage of another RSU, the user has to access the cloud tial things: A new reference model with insight analysis,” IEEE Access,
through the new RSU to retrieve the result by serializing session vol. 7, pp. 19653–19669, 2019.
information. Whereas in our approach, the cloud itself sends the [7] N. Fernando, S. W. Loke, and W. Rahayu, “Mobile cloud computing: A
survey,” Future Gener. Comput. Syst., vol. 29, no. 1, pp. 84–106, 2013.
result to the RSU, that is currently serving the device. Hence, [8] J. Zhang et al., “Energy-latency tradeoff for energy-aware offloading in
the delay and power consumption of the mobile device in the mobile edge computing networks,” IEEE Internet Things J., vol. 5,
proposed system are less than the existing system [16]. This is no. 4, pp. 2633–2645, Aug. 2018.
[9] K. Gai, M. Qiu, H. Zhao, L. Tao, and Z. Zong, “Dynamic energy-aware
observed that the proposed system reduces the delay and power cloudlet-based mobile cloud computing model for green computing,”
by approximately 55–60% and 57–74% respectively than the J. Netw. Comput. Appl., vol. 59, pp. 46–54, 2016.
existing system [16]. This is observed that for small as well as [10] Y. Zhu, Q. He, J. Liu, B. Li, and Y. Hu, “When crowd meets big video
data: Cloud-edge collaborative transcoding for personal livecast,” IEEE
large scale processing, our Mobi-IoST reduces the delay and Trans. Netw. Sci. Eng., vol. 7, no. 1, pp. 42–53, Jan.–Mar. 2020.
power consumption of the mobile device. [11] Q. Fan and N. Ansari, “Towards workload balancing in fog comput-
ing empowered IoT,” IEEE Trans. Netw. Sci. Eng., vol. 7, no. 1,
pp. 253–262, Jan.–Mar. 2020.
[12] C. Yang, Y. Liu, X. Chen, W. Zhong, and S. Xie, “Efficient mobility-
VI. CONCLUSION AND FUTURE WORK aware task offloading for vehicular edge computing networks,” IEEE
Seamless connectivity is a major challenge during data and Access, vol. 7, pp. 26652–26664, 2019.
[13] C. Wang, Y. Li, and D. Jin, “Mobility-assisted opportunistic computa-
computation offloading in any mobile network. The process- tion offloading,” IEEE Commun. Lett., vol. 18, no. 10, pp. 1779–1782,
ing of raw data collected using IoT devices and delivery of the Oct. 2014.

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.
 GHOSH et al.: MOBI-IOST: MOBILITY-AWARE FRAMEWORK FOR TIME-CRITICAL APPLICATIONS 2285

[14] F. Tang and H. Zhang, “Spatial task assignment based on information Shreya Ghosh received the B.Tech. degree from
gain in crowdsourcing,” IEEE Trans. Netw. Sci. Eng., vol. 7, no. 1, Indian Institute of Engineering Science and Technol-
pp. 139–152, Jan.–Mar. 2020. ogy, Shibpur, West Bengal, India, in 2015. She is
[15] H.-Q. Wu, L. Wang, and G. Xue, “Privacy-aware task allocation and currently working toward the Ph.D. degree with the
data aggregation in fog-assisted spatial crowdsourcing,” IEEE Trans. Department of Computer Science and Engineering,
Netw. Sci. Eng., vol. 7, no. 1, pp. 589–602, Jan.–Mar. 2020. IIT Kharagpur, Kharagpur, India. She is currently a
[16] A. Mukherjee, D. G. Roy, and D. De, “Mobility-aware task delegation Research Scholar with the Department of Computer
model in mobile cloud computing,” J. Supercomput., vol. 75, no. 1, Science and Engineering, IIT Kharagpur. Her current
pp. 314–339, 2019. research interests include spatial informatics, trajec-
[17] J. J.-C. Ying, W.-C. Lee, T.-C. Weng, and V. S. Tseng, “Semantic tra- tory data mining and cloud computing. She is the
jectory mining for location prediction,” in Proc. 19th ACM SIGSPATIAL recipient of the prestigious TCS fellowship.
Int. Conf. Adv. Geogr. Inf. Syst., ACM, 2011, pp. 34–43.
[18] Q. Liu, S. Wu, L. Wang, and T. Tan, “Predicting the next location: A
recurrent model with spatial and temporal contexts,” in Proc. 13th AAAI Anwesha Mukherjee received M.Tech. and Ph.D.
Conf. Artif. Intell., 2016, pp. 194–200. degrees from the Department of Computer Science
[19] A. Karatzoglou, N. Schnell, and M. Beigl, “A convolutional neural net- and Engineering, West Bengal University of Tech-
work approach for modeling semantic trajectories and predicting future nology, Kolkata, India, in 2011 and 2018, respec-
locations,” in Proc. Int. Conf. Artif. Neural Netw., Springer, 2018, tively. She is currently working as a Research
pp. 61–72.
Associate with the Computer Science Department,
[20] Y. Zheng, “Trajectory data mining: An overview,” ACM Trans. Intell.
IIT Kharagpur, West Bengal, India. Her research
Syst. Technol., vol. 6, no. 3, 2015, Art. no. 29. areas includes IoT, Fog computing, and mobile net-
[21] S. Ghosh, S. K. Ghosh, R. D. Das, and S. Winter, “Activity-based mobil- work. She was a recipient of the Young Scientist
ity profiling: A purely temporal modeling approach,” in Proc. 26th Int. Award from International Union of Radio Science, in
Conf. World Wide Web Companion, 2018, pp. 409–416. 2014, event held at Beijing, China.
[22] S. Ghosh and S. K. Ghosh, “Modeling of human movement behavioral
knowledge from GPS traces for categorizing mobile users,” in Proc.
26th Int. Conf. World Wide Web Companion, 2017, pp. 51–58.
[23] C. Song, Z. Qu, N. Blumm, and A.-L. Barabasi, “Limits of predictability Soumya K Ghosh is currently a currently a Professor
in human mobility,” Science, vol. 327, no. 5968, pp. 1018–1021, 2010. with the Department of Computer Science and Engi-
[24] S. Shekhar and S. Chawla, Spatial Databases: A Tour, vol. 2003. Upper neering, IIT Kharagpur, West Bengal, India. He was
Saddle River, NJ, USA: Prentice Hall, 2003. with the Indian Space Research Organization, Benga-
[25] C. E. White, D. Bernstein, and A. L. Kornhauser, “Some map matching luru, India. He has authored or coauthored more than
algorithms for personal navigation assistants,” Transp. Res. Part C: 200 research papers in reputed journals and confer-
Emerg. Technol., vol. 8, no. 1–6, pp. 91–108, 2000. ence proceedings. His current research interests
[26] A. Mukherjee, S. Bhattacherjee, S. Pal, and D. De, “Femtocell based include spatial data science, spatial web services, and
green power consumption methods for mobile network,” Comput. cloud computing.
Netw., vol. 57, no. 1, pp. 162–178, 2013.
[27] M. Vlachos, D. Gunopoulos, and G. Kollios, “Discovering similar multi-
dimensional trajectories,” in Proc. 18th Int. Conf. Data Eng., 2002,
pp. 0673.
[28] J. Yamato, J. Ohya, and K. Ishii, “Recognizing human action in time- Rajkumar Buyya is currently a Redmond Barry Dis-
sequential images using hidden Markov model,” in Proc. IEEE Comput. tinguished Professor and the Director of the Cloud
Soc. Conf. Comput. Vision Pattern Recognit., 1992, pp. 379–385. Computing and Distributed Systems (CLOUDS)
[29] M. Lv, L. Chen, and G. Chen, “Discovering personally semantic places Laboratory, University of Melbourne, Parkville, VIC,
from GPS trajectories,” in Proc. 21st ACM Int. Conf. Inf. Knowl. Australia. He is currently the founding CEO of Man-
Manag., ACM, 2012, pp. 1552–1556. jrasoft, a spin-off company of the University. He has
[30] C. Cheng, H. Yang, M. R. Lyu, and I. King, “Where you like to go next: authored more than 625 publications and seven text
Successive point-of-interest recommendation,” in Proc. 23rd Int. Joint books. He is one of the highly cited authors in com-
Conf. Artif. Intell., 2013, pp. 2605–2611. puter science and software engineering worldwide
(h-index = 128, g-index = 275, and more than 85 800
citations). He is recognized as a “Web of Science
Highly Cited Researcher” in 2016 and 2017 by Thomson Reuters, and Scopus
Researcher of the Year 2017 with Excellence in Innovative Research Award
by Elsevier for his outstanding contributions to Cloud computing.

Authorized licensed use limited to: ULAKBIM UASL - Lefke Avrupa Universitesi. Downloaded on November 17,2023 at 07:13:22 UTC from IEEE Xplore. Restrictions apply.