1

Mobi-IoST: Mobility-aware Cloud-Fog-Edge-IoTCollaborative Framework for

Time-Critical ApplicationsShreya Ghosh, Student Member, IEEE, Anwesha Mukherjee, Student Member, IEEE,

Soumya K. Ghosh, Member, IEEE and Rajkumar Buyya, Fellow, IEEE

Abstract—The design of mobility-aware framework for edge/fog computing for IoT systems with back-end cloud is gaining researchinterest. In this paper, a mobility-driven cloud-fog-edge collaborative real-time framework, Mobi-IoST, has been proposed, which hasIoT, Edge, Fog and Cloud layers and exploits the mobility dynamics of the moving agent. The IoT and edge devices are considered tobe the moving agents in a 2-D space, typically over the road-network. The framework analyses the spatio-temporal mobility data (GPSlogs) along with the other contextual information and deploys machine learning algorithm to predict the location of the moving agents(IoT and Edge devices) in real-time. The accumulated spatio-temporal traces from the moving agents are modelled using probabilisticgraphical model. The major features of the proposed framework are: (i) hierarchical processing of the information usingIoT-Edge-Fog-Cloud architecture to provide better QoS in real-time applications, (ii) uses mobility information for predicting nextlocation of the agents to deliver processed information, and (iii) efficiently handles delay and energy consumption. The performanceevaluations yield that the proposed mobility prediction algorithm has approximately 93% accuracy and reduced the delay and power byapproximately 23-26% and 37-41% respectively than compared to the existing mobility-aware task delegation system.

Index Terms—Cloud computing, Edge computing, Fog computing, Internet of Things (IoT), Mobility analytics, Spatio-temporal data.

F

1 INTRODUCTION

The advancements of Internet of Things (IoT) have man-ifested significant improvements on the quality of humanlives in varied aspects [1]. To facilitate real-time applications,high-end processing and storage units are required. Forcomputation and storage of these large volume of rawdata generated by IoT devices, cloud computing is themost significant infrastructure. However, the cloud-onlyset-up is not an energy-efficient and delay-aware solutionfor handling such a high volume of data. To address thisproblem, edge and fog computing have been introduced[2]. However, the seamless connectivity due to the mobilityof IoT devices is a crucial factor to process the data in theremote cloud servers. For time-critical applications such ashealth care, connection interruption and consequently theincrease in delay in delivering the processed information,result in poor Quality of Service (QoS). If the device getsdisconnected due to mobility, the delivery of the resultbecomes a challenge. In [3], the task delegation using pushnotification or serializing session information and applica-tion offloading using virtual machine migration have beendiscussed based on user mobility. Nevertheless, the device

• Shreya Ghosh and Anwesha Mukherjee are with the Department of Com-puter Science and Engineering, Indian Institute of Technology Kharag-pur, West Bengal, 721302, India.E-mail: shreya.cst@gmail.com, anwe-shamukherjee2011@gmail.com

• Soumya K. Ghosh is a Professor in the Department of Computer Scienceand Engineering, Indian Institute of Technology Kharagpur, West Bengal,721302, India.E-mail: skg@cse.iitkgp.ac.in

• Rajkumar Buyya is with the CLOUDS Laboratory, School of Computingand Information Systems, The University of Melbourne, VIC 3010,Australia.E-mail: rbuyya@unimelb.edu.au

has to receive the result from the cloud by serializing sessioninformation when gets connected [3]. Hence, the user has torequest explicitly for the result. For critical real-time appli-cations, this may not be a good solution. This necessitatesa hierarchical infrastructure, where each layer (IoT, edge,fog or cloud) either accumulates, stores and processes theinformation for reducing the delay.

On the other side, movement traces, i.e., time-stampedlocation information of moving agents (say, mobile-users orclient) are accumulated on a large scale from GPS-enabledsmart phones or IoT devices. This spatio-temporal move-ment information opens up diverse opportunities to explorethe intent of movement [4] and thus fostering varied loca-tion based services, namely, efficient package delivery [5],traffic resource management etc. Internet of Spatial Things(IoST) brings IoT in the spatial context [6]. As discussedbefore, mobility or continuous change of locations of users isa challenging issue in task delegation or data offloading.However, analysing these mobility information helps toexplore the intent of the move and subsequently extracts thefrequent movement path of a user in different contexts. If theprobable location sequences of an agent in the near futurecan be predicted from the historical mobility information,then an effective and delay-aware solution for a time-criticalapplication can be provided.

To address the above-mentioned challenges, in thiswork, we propose a Cloud-fog-edge based collaborativeframework for the processing of IoT data and deliveringthe result based on mobility analysis to reduce the delay. Wehave considered a hierarchical mobility-based infrastructurecomposed of four layers: IoT layer, edge layer, fog layer,and cloud layer. Nowadays smart phone has become a

2

popular medium for ubiquitous Internet access and varieduser-specific IoT applications are accessible through smartphones. These mobile devices serve as edge devices andconstantly change the locations. The users of our systemutilize these time-critical applications while traveling across.The edge layer contains such edge devices i.e. mobile de-vices. The fog layer contains the fog devices such as RSUswhich are large cell base stations. While the IoT and the edgedevices change their locations, the RSU and the cloud datacenters of the framework have static locations. The raw datagenerated in the IoT layer is sent to the edge layer, which isconnected with the fog layer. The fog layer is connected withthe cloud layer where high-end processing and mobilityanalysis tasks are performed.

Fig. 1: Mobi-IoST for health care application(a): Ambulance sends health data generated from IoT devicesto RSU.(b): RSU sends the result to cloud along with the locationinformation(c): Cloud predicts the nearby health care centre, shortest pathand helps to actuate traffic signal

1.1 Motivating Scenario

We have considered a well-known time-critical application,health care, where the proposed framework Mobi-IoST canbe utilized. The pictorial representation of this use-case isshown in Fig. 1.

Suppose a patient, travelling in an ambulance (am),needs continuous monitoring of her/his vital health param-eters. The health parameters such as blood-pressure, pulse-rate, body-temperature etc. are collected using IoT devicesand the raw data are sent to the RSU through a clientapplication. The RSU processes the information and sendsthe current status as normal/abnormal to the client-app. Ifan abnormality is detected, the RSU sends the data to thecloud to find out the nearest health facility. Given the currentlocation and health-data feed from the RSU, the cloud cansuggest the nearby hospital. On the other side, based on theroute followed by the ambulance, the probable health centrealso gets notified. Further, the mobility analysis moduleof cloud can help to reduce the commuting time of theambulance by predicting less congested path in the road-network. This can be achieved when cloud analyses thetraffic states (congestion, traffic breakdown etc.) of the roadsand notifies the RSUs of the path. The RSU will work

as a fog device and the respective RSUs can actuate thesignal synchronizing mechanism such that am can reachavoiding the congested route as well as without waiting inthe traffic signals. Although the scenario is motivated bythe dysfunctional public health system and limited accessto improved transportation and medical care in the ruralareas, specifically in developing countries, such as India,Mobi-IoST is beneficial in other time-critical applications aswell. For instance, in the time of emergency, a police-vehicleor a fire extinguisher car needs to commute with a minimaldelay avoiding the congested regions of a city. Mobi-IoSTpredicts the less congested route by analyzing the trafficstates in real-time and notifies the RSUs of the route. TheseRSUs actuate the signal synchronizing mechanism such thatthe vehicle can reach the destination avoiding the congestedroute as well as without waiting in the traffic signals. Thehierarchical placement of IoT, Edge, Fog devices and cloudservers in Mobi-IoST framework facilitates an effective anddelay-aware solution for several time-critical applications.We believe that Mobi-IoST will act as a foundation ofmobility aware network resource management for variedlocation-based service planning in real-time.

1.2 Contributions

The focus of our work is to develop a Cloud-fog-edgecollaborative framework which facilitates real-time IoT in-formation processing and delivery of results based on themobility information analytics. The key contributions of thispaper can be summarized as follows:

• Mobi-IoST (Mobility-aware Internet of SpatialThings) is designed for information processing anddelivering result based on the prediction of the user’scurrent location. The framework exploits the mobil-ity knowledge of the agents to predict the probableuser location and delivery of processed informationat low delay and low power consumption of the user-device.

• A novel mobility modelling network has been pro-posed to explore the movement patterns of the user.The huge amount of spatio-temporal trajectory datais stored efficiently along with other contextual infor-mation in the cloud data centre.

• A real-time mobility prediction module has beendesigned to predict the location sequences of the usereffectively.

• The experimental results demonstrate that the pro-posed system has outperformed other existing ap-proaches in accuracy and takes much less time tolearn the patterns.

• The simulation results demonstrate that the pro-posed framework reduces the delay in deliveringinformation and power consumption of mobile de-vice (user-device) compared to the existing mobility-aware task delegation approach.

To the best of our knowledge, this work is the first attempt toutilize the movement knowledge to enhance the QoS for fa-cilitating time-critical IoT applications. The rest of the paperis structured as follows. Section 2 briefs the existing workin related areas. We propose our framework, Mobi-IoST in

3

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

section 3 and discuss several modules of the framework.The delay and power consumption models are discussedin section 4. Section 5 demonstrates the experimental andsimulation results. The paper is concluded in section 6 alongwith future directions.

2 RELATED WORK

The IoT refers to the connection of embedded devices withinan existing Internet infrastructure where the devices areuniquely identified and the computing environment is cre-ated [1]. The sensor nodes are usually used as IoT devices,which collect the object status. The raw data collected byIoT devices are processed inside the cloud servers. How-ever, storing and processing of the raw data inside theremote cloud enhances the delay and energy consumption.To overcome this, fog computing has been introduced [2].Nowadays IoT and fog become interrelated to each other[2]. The raw data of IoT devices are processed inside thefog device instead of the remote cloud to reduce the delayand energy consumption. However, during data processingconnection interruption becomes a challenge if the client isa mobile device. IoT has several sub-domains depending onits applications e.g. Internet of Multimedia Things (IoMT),Internet of Health Things (IoHT), Internet of Vehicles (IoV)etc [6]. IoST is a new sub-domain of IoT, which focuseson spatial data management [6]. IoST refers to “ubiquitousand embedded computing devices that transmit and receiveinformation so often includes numerical values about aphysical object that can be represented in a geographic co-ordinate system for geospatial interoperability requirementsover networks”[6]. In fog-based IoT, the switch, routers etc.work as fog devices for faster processing of the raw datacollected using IoT devices. The mobile device that usuallyworks as an edge device, is a connector between the IoTdevices and the network. Though nowadays different appsare present in smart phones, resource hindrance is a majordifficulty for these handheld devices. Battery life of the de-vice, memory, storage become major constraints. Therefore,the cloud servers have to be used by mobile devices to

store their data [7]. The mobile devices also offload heavycomputations inside the cloud in case of resource limitationand saving battery life. However, for small amount of dataprocessing the use of remote cloud increases delay andpower consumption of the mobile device. But the device isunable to do the processing due to resource limitation. En-ergy and latency in offloading has been focused on severalexisting approaches [8], [9]. The use of cloudlet for reduc-ing energy consumption in the multi-cloudlet scenario hasbeen highlighted in [9]. Fog computing has also providedsolutions for reducing delay and energy in the processing ofIoT data [2]. In fog computing, a hierarchical architectureis followed, where the intermediate devices between theend node and cloud servers participate in data processing,and these nodes are called fog devices [2]. The edge devicesallow users to connect with the network and transfer dataaccordingly to a network which is external to the user.For transcoding massive amount of video at scale, a cloudand edge computing based collaborative system has beenproposed in [10]. For balancing the traffic and computingload, a method has been discussed in [11], where the IoTdevices are allocated to the base station or fog nodes toreduce the latency.

Network connectivity is a challenge in vehicular net-work [12]. For task offloading in such networks, edge com-puting has been used in [12]. The mobile edge computingservers are deployed inside the road side access pointsin that case. These servers are used for offloading tasks.However, the use of access points may not be energy-efficient if exhaustive computations have to be performedand there are a large number of users. Based on usermobility, an opportunistic computation offloading methodhas been discussed in [13]. Based on information gain,task allocation in spatial crowdsourcing has been discussedin [14]. A fog based architecture of spatial crowdsourc-ing has been proposed in [15]. In this work, the authorshave focused on privacy-aware task allocation and dataaggregation. In [3], task offloading to cloud and deliveryof result based on serialization of session information hasbeen discussed. However, the user if gets disconnected has

4

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

to retrieve the result by sending a request to the cloud. Forthe cloudlet based scenario, virtual machine migration hasbeen discussed in [3]. Nevertheless, the VM migration has totake place from one cloudlet to another until the processinggets finished and the result is delivered to the user. Thisis expensive as well as increases the energy consumption ofthe cloudlet. Hence, there should be a mechanism where theuser mobility will be predicted and result will be deliveredat the optimum delay and power consumption of the mobiledevice.

Given the abundance of mobility traces (GPS log) ofindividuals, there are several research initiatives to extractknowledge or meaningful information (i.e., making sense ofraw GPS log) from the huge amount of trajectory traces.Several works are reported to predict next location frommovement traces such as GPS log, check-in data or socialnetwork information [16], [17], [18]. There are challeng-ing applications, namely, urban land-use classification fromtaxi-traces [19], categorizing users in an academic campus[20] or catching pick-pockets from large-scale transit records[21]. It is well known that human movement traces followspatio-temporal regularity. In this regard, Song et al. [22]provide a high degree of spatio-temporal uniformity bymining movement traces of 50,000 people for a period ofthree months. All of these studies depict that since peoplefollow some spatio-temporal regularity in their movementhistory, therefore an appropriate and effective mobility pat-tern modelling can help to facilitate several location-awareservices.

To this end, the Mobi-IoST framework aims to delivermobility-driven efficient data processing in cloud-fog-edgebased IoT setup to facilitate intelligent decision making inreal-time. One of the major aspects of the proposed frame-work is mobility-aware service provisioning, which helps toreduce the delay, power consumption of the user-device andas well as facilitating intelligent recommendations basedon the present location of the user-device. For example,

recommending the nearest health-care center based on thehealth condition of the patient (say, cardiac problem) and theavailable facilities of the health-care centers. Furthermore,the RSUs can help to synchronize the signaling mechanismbased on the traffic states of a road network (Fig. 1) . Tothe best of our knowledge, there are several research effortsin the domain of mobile big data mining and fog basedIoT, but is somewhat fragmented and no other existingworks have clearly depicted the significance of mobility-aware service provisioning framework in fog based IoT. InMobi-IoST, the movement pattern modelling and locationprediction approaches are novel propositions which deliverresult in real-time. Moreover, the experimental observationsand performance analysis show the effectiveness of Mobi-IoST in terms of accuracy, delay and power consumption. Insummary, designing and deploying an end-to-end mobilitydriven framework for efficient data processing in IoT setupis a challenging issue in the present era.

3 MOBI-IOST FRAMEWORK

The hierarchical structure of Mobi-IoST is represented inFig. 2. Fig. 3 depicts the overall flow of the framework,Mobi-IoST. IoST or Internet of Spatial Things deals with IoTdata along with spatial perspective. As depicted in Fig. 2,in the bottom layer several IoT sensors such as accelerom-eter, GPS, temperature, blood-pressure, proximity sensorscapture application specific data. These IoT sensors areeither present within the edge devices or connected withthe edge devices, namely, mobile phone, vehicles, whichchange their locations. When any of these edge devicesneeds assistance, it contacts the nearest RSU. In this work,RSU is used as fog device and it is capable of small scaleprocessing. If the processing is beyond the computationalcapability of the RSU, then it delegates the task to the cloud.The top layer of the hierarchical structure consists of cloudservers, which store spatial data, specifically, mobility traces,location-specific information, city-structure (POI placements

5

and other contextual information). The cloud processingunit executes the task and sends the result to the RSU, wherethe tasks are application-specific. For example, in case ofan emergency, an ambulance needs to reach the destinationwithout any delay. The cloud processing unit extracts anddiscovers the present traffic status and predicts the pathwith minimum commuting time. As the cloud storage unithas all the RSU information, it notifies all the RSUs withinthe predicted path for signaling and actuating such that theambulance does not face any traffic congestion. Further,for any personalized recommendation, a mobile user canalways send a request to the nearest RSU. Suppose, auser captures her health-related data, namely, temperature,blood-pressure and weight using the IoT devices and sendsa request to the RSU through the mobile device for pre-dicting the current health status. Furthermore, the resourcescan be efficiently managed by this framework: movementanalysis module can predict variation of travel demandapriori and notify the RSUs accordingly, while the RSUscan decide about the dissemination of resources (traffic ornetwork) efficiently.

The major modules of the framework are: (i) movementpattern modelling, (ii) predicting next location sequences,(iii) delivery of result after processing in a timely manner.Finally, the experimental and simulation results yield theeffectiveness of the framework. The detailed approaches arediscussed in the remainder of the paper.

3.1 Exploring Movement Semantics from TrajectoryTracesThis section presents the methodology to model movementpatterns and predicts the next location sequences efficientlyand timely manner. Location prediction of moving agents,such as, people, vehicles is a challenging task for variedlocation-based services [23]. Specifically, in our work, loca-tion prediction helps to locate the moving agent’s locationsin near future and subsequently data is sent to the appro-priate RSU. Whenever a mobile device gets connected toa RSU, the GPS log of the mobile device is extracted andstored in the cloud dynamically.

It may be noted that location prediction depends onseveral factors, namely, day of the week, time-slot of a dayand road-structure. The first step of movement behaviourmodelling is to find out the frequent pattern followed bythe users in varied contexts. For example, the path followedby an individual differs significantly in weekends comparedto his/her weekdays’ trajectory signature. Moreover humanmovements follow some intent [4] and extracting the pur-pose behind any move is the fundamental step to predictnext location effectively. Few preliminary concepts whichare used in this paper are defined as follows:

• GPS log (G): GPS log is the collection of time-stamped latitude, longitude information. The GPStrajectory or trace is formed by connecting the lo-cation information on increasing time-ordering.Traj(p1, . . . , pn) :< p1(lat1, lon1), t1 >→<p2(lat2, lon2), t2 > · · · →< pn(latn, lonn), tn >,where t1 < t2 < · · · < tn

• Stay-Point (S): Stay-point of a trajectory is definedas a location (typically, polygon), where the moving

agent stops for a time-value δt and δt > tthresh.Here, polygon is a data-type of spatial data [24] andthe area of the polygon is less than arear , tthreshis the time-threshold for detecting stay-points fromthe trajectory. In our analysis, we have consideredthe parameter values as, tthresh = 12mins andarear = 2km2.

• POI and Geo-tagged Trajectory: Point-of-interest (POI)of a GPS location denotes the nearby landmark of alocation, such as, residential area, supermarket etc.We have followed the POItaxonomy

1 to extract suchPOI information using Google Place API. Geotaggedtrajectory is generated by appending the geo-taggedinformation of the stay-points within the trajectory.

• Trajectory window (TrajW ): Trajectory windowstores the location sequence information betweentwo such stay-points in an uniform sampling rate.

Augmenting Semantic Information with GPS log:Human movement semantics can be analysed if additionalinformation such as, POI, road-network structure and stay-point information are appended with the raw GPS traces.

- Road network of the study region is extracted fromOpenStreetMap (OSM) 2. The road network is rep-resented by a directed graph R = (V,E), wheree ⊆ |E| denotes the road-segments of the regionand v ⊆ |V | is the intersection points of such roadsegments. Map-matching algorithm [25] has beendeployed, which considers both geometric and topo-logical structure of the road-network to associate theroad-segments along with the trajectory traces.

- Each stay-point of the trajectory is geo-tagged withthe nearby POI location. Here, we have implementedthe iterative reverse geo-coding technique to extractnearest landmark of the stay-point.

After the addition of semantic information with the rawtraces, a trajectory trace takes the form:< pa, ta, Residential >, TrajW [(pi, ti, ex), (pi+1, ti +δt, ex), (pi+2, ti + 2× δt, ex), . . . ],< pb, tb, SuperMarket >, TrajW [(pj , tj , ey), (pj+1, tj +δt, ey), (pj+2, tj + 2× δt, ey), . . . ],< pc, tc, Residential >.Here, pa, pb and pc are three stay-points with geo-tagged in-formation residential building and supermarket. TrajW storesthe route information followed by the trajectory, whereex, ey are the road-segments of the road-network of thestudy region.

Processing of Large Mobility Datasets in CloudWith the advances in sensor technologies and the prolifera-tion of smartphones, a huge amount of mobility traces aregenerated by moving agents. One of the major challengesis to analyze the vast amount of data due to computationalcomplexity and storage limitations. To this end, we proposeto migrate the computation of mobility analysis and storageof movement traces in the cloud for faster response. It maybe noted that the locations and coverage areas of the RSUs

1. https://developer.foursquare.com/docs/resources/categories/2. OpenStreetMap: https://www.openstreetmap.org

6

need to be maintained in the cloud storage such that it canpredict the next location of the moving agent to determinethe appropriate RSU, which will serve the agent at thattime. Here, large cell base stations [26] are the RSUs. Themacrocell base station is referred to as macro RSU andmicrocell base station is referred to as micro RSU. Thecoverage area of macro RSU and micro RSU are 1-20km and200m-1km respectively [26]. The framework is implementedin Google Cloud Platform (GCP) by utilizing several storageand computational components of GCP. In our framework,cloud storage is of four types:

- Grid based storage of the study region: The studyregion is segmented into uniform hexagonal gridsand information, such as, road structure or POIs,RSUs are associated with each such grids. Our propo-sition is to segment the spatial region into grids suchthat each grid encloses the coverage area of at leastone micro RSU.The grid-segmentation process initiates with the lo-cation of one RSU. Suppose, the location of the RSU(say, RSUi) is p = (x, y) and the length of the side ofthe hexagon (gi) is a = 8m. An iterative process is de-ployed until the complete study region is segmentedwith hexagonal grids. In the first iteration, centerpoints of the 6 neighbouring grids of gi are calculatedand subsequently, the neighbouring hexagonal gridsare constructed.In the next step, Geohash code of all hexagonal gridsare computed. Geohash code of the grids representthe spatial location on the earth surface using uniquealphanumeric strings. Cloud Spanner of GCP is usedto store these information which supports horizontalscalability.

- Road network information storage: This modulestores the road network information, namely, connec-tions among different road-segments and road-type(highway, lane etc.). The information is stored in anadjacency matrix format, where each vertex maintainsa list of outgoing edges (outdegree of the vertex). Thedata-type of the list is polyline, which is a spatial-datatype [24] and represents the road-segments on themap.

- RSU information storage: It stores the list of RSUsalong with the unique-id, coverage area, location (lat-itude, longitude) and other information such as, type(micro RSU or macro RSU) etc. Cloud BigQuery ofGCP is utilized to store the road network informationand RSU information as well.

- Frequent path storage: The frequent path followedby individual moving agents are extracted and mod-elled in our work. Details are presented in section3.1.1. Cloud Bigtable of GCP is utilized to store theroad network information.

The computational cost of the mobility traces is hugesince it deals with time-series data with very high samplingrate. The key challenge is to reduce the processing time ofthe location prediction, and therefore, an efficient schemeis required. Here, we have deployed a hash-based indexingscheme, where nearby locations are stored in the subsequentbuckets of the hash-table.

3.1.1 Movement behaviour modellingIn this section, we discuss how movement behaviourof users can be modelled to explore the frequent pathsfollowed by them in different contexts. The process ofsemantic enrichment of GPS log of users has alreadydiscussed in section 3.1. Here, we propose User movementgraph, a multi-layer graphical model to model the users’movement patterns from the spatio-temporal context.The objective to use the multi-layer network is thathuman movement patterns typically depend on temporalvariations (weekdays or weekends, morning or evening),road networks and stay-points. All of these informationneed to be encoded and interconnections of the informationcannot be properly captured in a single layer.

User movement graph (MG): User movement graph isdefined as MG = (N,L, la), where N denotes the nodes,L denotes the links and label is represented by la. The usermovement graph has four labels:

• Road network: The layer 1 consists of road networkinformation, where nodes are intersection points ofroad-segments.

• RSU network: The RSU information (location andcoverage area) is stored in layer 2.

• Stay-point information: The stay-point informationincluding location and type of stay-point are storedin layer 3.

• Frequent path: The movement paths frequently fol-lowed by the user is stored in layer 4.

It may be noted that each layer is interconnected with eachother. As the construction of layer 1, layer 2 and layer 3 arestraight forward, we discuss the frequent pattern miningprocess of layer 4 in detail.

The frequent path network of layer 4 is represented byprobabilistic graphical model or Dynamic Bayesian networkFPN(V,E,Υ) where V is the set of stay-points, E denotesthe direction of visit among different stay-points and Υ isthe network quantify parameter. The key intuitive to uti-lize probabilistic graphical model is that the visit sequenceof a person somewhat follows conditional dependency. Insimple words, whether a person will visit location l1 orl2 at t + 1, depends on her present stay-point at t. In thiswork, we have considered both spatial location and temporalspan of a visit-sequence to model FPN of user movementgraph. Each node (stay-points: v ⊆ V ) of the network isconditionally independent of its non-descendants given itsparent node (Pa(v)). Suppose, a visit-sequence is given asV = (V1, . . . , VN ), the probability distribution is computedas follows:

P (V ) =

N∏i=1

P (Vi|Pa(Vi)) (1)

FPN captures the dynamic nature of the mobility infor-mation by representing multiple copies of the spatial-information, one for each time-slice Vt = (V1,t . . . , Vd,t).Subsequently, the transition distribution from one stateto other (P (Vt+1|Vt)) is computed from two time-sliceBayesian network. The spatial location information (Vt) istypically divided into two sets, namely, unobserved statevariables (St) and the observed state variables ( Lt, in our

7

case, location information from RSUs). The joint probabil-ity distribution is calculated by unrolling two time-sliceBayesian networks:

P (S0, . . . , ST , L0, . . . , LT ) =

P (S0)P (L0|S0)

T∏t=1

P (St|St−1)P (Lt|St)(2)

It may be noted that we have represented the stay-points asgrid-location and transition from one state to another statesignifies that the agent is moving from one grid to another.

Next, we deploy a spatio-temporal trajectory clustering(TrajCS) on FPN which captures the signature or fre-quently visited paths of the individual. The process followsa hierarchical top-down approach, and based on the dis-tance measure new clusters are generated and appendedin the list. The trajectory clustering distance measure iscomputed as follows:TrajCS(Si, Sj) =

0 if(i == 0)or(j == 0)

TrajCS(Si−1, Sj−1) if((Si == Sj)+C ×Min(TScorei , TScorej ) and(si+1) 6= (sj+1))MAX(TrajCS(Si−1, Sj),

T rajCS(Si, Sj−1)) if(si 6= sj)

(3)

where Si represents a set of locations, si denotes one GPSpoint of the set Si and C is the parameter to augment theprobability of the path taken. Tscore computes the temporalsimilarity between two different stay-points of the trajectoryand TrajCS method is recursively called for extracting thesignature pattern. The proposed distance measure appendstemporal information with the conventional LCSS cluster-ing method [27]. Algorithm 1 describes the basic steps ofgenerating FPN from the trajectory traces of agents.

This section describes how users’ frequent movementpatterns are extracted and stored along with other contex-tual information. Furthermore, the trajectory clustering andmulti-layer graphical model help to effectively model themovement behaviour of agents in the cloud server.

3.1.2 Next location sequence prediction from mobility infor-mationThis helps to calculate the probable path visit by the userapriori. The location prediction task is formulated as fol-lows: Given the historical observations of an moving agent mand the current location L at time t, predict the agent’s anticipatedlocation sequences (S) following δ time-instances

Typically, the task is formulated as information retrievaltask considering the fact that people’s movement patternsfollow spatio-temporal regularity and effective movementbehaviour modelling leads to accurate location prediction.Here, we have deployed Hidden Markov model (HMM) (sayχ) based prediction technique with two kinds of stochasticvariables, state variables (hidden) and observable variables. Eachindividual’s movement is modelled as kth order Markovchain and the transition from one place to another place ismodelled based on MG and χ.

Algorithm 2 describes the basic steps of location pre-diction. The first step map locates the current location (gridlocation) of the moving agent and then the model predicts

the sequences of locations based on the context and finally,update process updates the result based on the current inputfrom the mobile device. It may be noted that the order (k) ofthe markov chain is dependent on the user’s frequent move-ment pattern and extracted from FPN of user movementgraph (MG).

P (si|si−1, si−2, . . . , s1) = P (si|si−1, . . . , si−k) (4)

In the next step, forward algorithm [28] is deployed to findout the sequences of stay-points, given as:

P (Lk|χ) =

seqmax∑i=1

[

k∏j=1

P (L(j)|Si(j))

∗ P (Si(j)|Si(j − 1), Si(j − 2), ..., 1)]

(5)

where, seqmax and Si(j) represent the maximum num-ber of hidden state sequences and hidden states. Here,model is represented as k-order markov chain where the nextlocation depends on k recent observations. Next, a variantof verterbi algorithm using time-relationships is deployed todiscover the possible sequences of states. The transition andemission probabilities of χ are computed by adjusting themodel parameters. An iterative version of forward backwardalgorithm is implemented to produce the sequences effec-tively.

Three types of location prediction tasks have been car-ried out in this work: (i) location sequence prediction in aspecific time-threshold, (ii) predicting appropriate POI (say,health-care center) and the path and finally, (iii) given thedestination and present location of the agent predicting thepath with less commuting time based on the traffic statesof the road-network. The location sequence prediction inspecific time-thresholds is computed directly from χ onMG. The POI and path prediction is carried out by over-lapping the road-network structure (layer 1 of MG) andfrequent path pattern (layer 4 of MG). Finally, given sourceand destination, the markov model is used along with thetraffic-state of the region, where s1 and sn are specified.It may be noted that our algorithm (Algorithm 2) is self-adaptive, i.e., the update function (also, refer ’Update’ arrowof Fig. 3) changes the modelling algorithm in case any of theprediction result fails.

3.2 Delivery of Result After Data Processing

The IoT devices are connected with the edge device, e.g. thesensors within the mobile device. With the increasing avail-ability of smartphones, we have considered mobile devicesas the edge devices. The mobile device will process rawdata received from the IoT devices. If the mobile device isable to perform the processing, it does the same by workingas an edge device and generates the result. Otherwise, themobile device sends the data to the RSU, which will act asa fog device. Each RSU maintains a look-up table, whichholds the mobile device IDs present under its coverage. TheInternational Mobile Equipment Identity (IMEI) number isconsidered as the mobile device ID. The RSU after receivingthe raw data from the mobile device, checks its current loadand ability to process the data. In this regard, two casesappear as follows:

8

Algorithm 1 : Frequent path mining - A trajectory clustering approach

Input: Set of trajectory T , stay-points SOutput: Frequent path network < FPN(V,E,Υ) > . Frequent path network for each individual

1: clus, S, V,E ← NULL;2: for each trajectoty window tr ∈ T do3: for each unvisited stay − point s ∈ S do4: V.append(s) . Create new node5: visited← s6: Υ : CPT ← Create ConditionalProbabilitytable(s)7: t← extractTemporal(s) . Temporal information of the stay-point8: E.append(genEdge(S, t)) . genEdge creates directed edge based on frequency of the visit and temporal information9: end for

10: end for11: for each path trp ∈ FPN do . Path represents the successive sequence of stay-points12: Ne← ExtractTrajW (trp, FPN) . Extract trajectory-windows within spatio-temporal locality13: D ← computeLCSS(Ne, trp)14: if D > thresh then15: Ignore Ne16: end if17: if D < thresh then18: Create new cluster clust19: clustt.append(Ne, trp) . Appending new cluster in the list20: visited← Ne21: Modify CPT (clustt) . Modify the CPT of all nodes in clustt calculating the frequency of visit22: end if23: end for

Algorithm 2 : Location prediction - map and update process

Input: User movement graph MG, Present location s, Trajectory log TOutput: < s′, Edge− list > . sequence of next location sequences

1: function MAPPER(s,MG, T ) :2: j ← geo− hascode(s) . extract the grids where trajectory points placed3: E′ ← extract pattern(MG, j) . extract the frequent trajectory patterns within the grid4: for all ti ∈ T do5: L← predictLoc[χ(ti, s)] . HMM based prediction6: p← ComputeProb(s, arraylist[L, t]) . Compute transition probability for all patterns in partcular time-stamp7: dist← ComputeTrajCS(E′, ti) . Compute the trajCS distance for each such patterns8: s′ ← SORT (arraylist[p], arraylist[dist]) . Predicted location with maximum probability9: end for

10: Print < s′ : arraylist(E′0) >

11: function Update(s, arraylist[s′]) : . If sudden change in mobility pattern observed12: for all ti do13: for all ea ∈ arraylist[s′] do14: Append trajectory window Tre containing ea in FPN15: ModifyCPT (ea, ti) . Modify CPT of all nodes with outgoing/incoming edge ea16: ModifyProb(ea, ti) . Modify transition matrix of all sequences containing ea17: Mapper(s,MG, Tre)18: end for19: end for

• If the RSU’s current load is equals to the maximumload it can handle or an exhaustive computation isrequired to perform which is beyond the capabilityof the RSU, it forwards the data to the cloud alongwith the device ID and request ID. After processingthe data, the cloud finds the current location of thedevice based on the user’s geo-location information(see section 3.1). Based on the current location of thedevice, the cloud identifies the RSU under which themobile device is currently located. The cloud sendsthe result to the RSU along with the device ID andrequest ID. The RSU forwards the result to the mobiledevice.

• If the RSU is able to process the raw data and itscurrent load is less than the maximum load it can

handle, the RSU processes the data and sends theresult to the mobile device. However, as the device isin mobility, it may be possible that the RSU finishesprocessing and the mobile device moves away. Insuch a case, the RSU sends the result to the cloudalong with the request ID and device ID. The currentlocation of the device is predicted by the cloud basedon the user’s geo-location information. Based on thecurrent location of the device, the cloud identifiesthe RSU under which the mobile device is currentlylocated. The cloud sends the result to the RSU alongwith the device ID and request ID. The RSU forwardsthe result to the mobile device.

In our approach as the mobility information is updateddynamically, the probability of the presence of the mobile

9

TABLE 1: Symbols used in power and delay calculation

Symbol Definitionu Velocity of moving agentDc Amount of data collected and transmittedDr Amount of result data received

Uptmr Data transmission rate in uplink betweenmobile device and RSU

Dwtmr Data transmission rate in downlink betweenmobile device and RSU

Uptrc Data transmission rate in uplink betweenRSU and cloud

Dwtrc Data transmission rate in downlink betweenRSU and cloud

fmr Uplink data failure rate betweenmobile device and RSU

frm Downlink data failure rate betweenmobile device and RSU

frc Uplink data failure rate between RSU and cloudfcr Downlink data failure rate between RSU and clouddpr Data amount processed per unit time by the RSUdpc Data amount processed per unit time by the cloudTm Delay to determine the current location of the devicePa Power consumption of mobile device in active modePi Power consumption of mobile device in idle mode

device under the predicted RSU is high. However, if thedevice losses connection with the network for a long du-ration, there is a probability that the mobile is not locatedunder the coverage of the predicted RSU. In such cases,after receiving the result from the cloud, the predictedRSU sends feedback to the cloud that the mobile device isnot present under its coverage. By this time, if the mobiledevice gets connected with a RSU, it will send a request forthe result with the request ID to the RSU. The RSU thenforwards it to the cloud. The cloud then sends the result tothe mobile device through the RSU and updates the usermobility information accordingly. Algorithm 3 summarizesthe steps of the working model of the proposed frameworkMobi-IoST. As we observe in the proposed system the cloudafter analysing the user mobility information, finds out theprobable RSU currently serving the device and accordinglydelivers the result. Hence, intelligent decision making isperformed by the cloud, which makes the system efficient.

4 DELAY AND POWER CONSUMPTION IN MOBI-IOSTThe mobile device transmits the data to the RSU with whichit is currently connected and requests for providing theresult after processing. According to our strategy, either RSUor cloud performs data processing based on the complexityof the computation required for processing the data and thecurrent load of the RSU. Here, two cases are considered asdiscussed previously:

• Information processing inside the RSU• Information processing inside the cloud

The symbols used in the delay and power calculation of theproposed model, are defined in TABLE 1.

4.1 Delay model for Information Processing in RSUThe uplink data transmission delay between mobile deviceand RSU is given as:

Tmr = (1 + fmr) ∗ (Dc/Uptmr) (6)

The downlink data transmission delay between mobile de-vice and RSU is given as:

Trm = (1 + frm) ∗ (Dr/Dwtmr) (7)

The total data transmission delay between mobile deviceand RSU is given as:

Tt = Tmr + Trm (8)

The data processing delay inside the RSU is given as:

Tpr = Dc/dpr (9)

The total delay for data transmission and processing is:

Ttot = Tt + Tpr (10)

The mobile device while located at point i transmits the datato the RSU and requests for processing. Let the radius of theRSU’s coverage area is R and the last visited point of themobile device inside the RSU is k. The delay in movementfrom point i to k is given as:

Tik =

k−1∑i=1

((Di(i+1))/ui) (11)

where ui is the velocity of the mobile device at locationpoint i and Di(i+1) is the distance between two consecutivelocation points i and i + 1. If Tik > Ttot, the mobile deviceis still inside the coverage of the RSU. Hence, the RSU willdeliver the result to the mobile device. Hence, the round-tripdelay is given as:

Tdel1 = Ttot (12)

The power consumption of the mobile device during thisperiod is given as:

Pdel1 = Tt ∗ Pa + Tpr ∗ Pi (13)

As the RSU performs data processing instead of the cloud,the transmission delay is reduced. Hence, the delay indelivering the result is reduced in our system. Accordingly,the power consumption of the mobile device is also reduced.Else if Tik ≤ Ttot, the mobile device has moved to thecoverage of another RSU. Hence, the previous RSU willforward the result to the cloud along with the request IDand the mobile device ID. The cloud contains the mobil-ity information of the mobile devices. Using the locationprediction strategy described in the section 3.1.2, the cloudwill find out the current probable location of the mobiledevice. Let t is the time instant when the mobile device isat location i. The cloud finds out the location point visitedby the mobile device at time instant (t+ Ttot), and the RSUwhich is currently serving the device. The cloud delivers theresult to the selected RSU, which forwards the result to themobile device. In this case, the round-trip delay is:

Tdel21 = Ttot + (Dr/Uptrc)(1 + frc)

+Tm + (Dr/Dwtrc)(1 + fcr)(14)

where Tm is the delay for determining the current locationof the device and correspondingly the RSU currently servingthe device, based on the mobility information of the user.

10

Algorithm 3 : Working Model of Mobi-IoST

Input: Raw data received from IoT devicesOutput: Result after processing the raw data

1: mobile device receives raw data from IoT devices2: if mobile device is able to process the data then3: mobile device works as edge device and processes the data4: else5: mobile device forwards the data to the fog device RSU6: if current load of the RSU < maximum load the RSU can have then7: go to step 118: else9: go to step 34

10: end if11: if RSU is able to process the data then12: RSU processes the data13: if the mobile device is still connected then14: RSU delivers result to the device15: else16: RSU forwards result to the cloud along with device ID and request ID17: cloud predicts current location of the mobile device from the mobility information using Algorithm 1 and 218: cloud identifies the RSU serving the predicted location19: cloud forwards the result to the predicted RSU along with the device ID and request ID20: if the mobile device is connected with the predicted RSU then21: RSU sends the result to the mobile device22: else23: RSU sends a feedback to the cloud that the mobile device is not present in its coverage24: cloud after receiving the feedback stores the result25: if the mobile device gets connected with a RSU then26: mobile device requests for the result to the RSU with the request ID27: RSU forwards the request to the cloud28: cloud sends the result to the RSU and updates the mobility information29: RSU sends the result to the mobile device30: end if31: end if32: end if33: else34: RSU sends the raw data along with device ID and request ID to the cloud35: cloud processes the data36: go to step 1737: end if38: end if

The power consumption of the mobile device during thisperiod is given as:

Pdel21 = Tt ∗ Pa + (Tpr + (Dr/Uptrc)(1 + frc)

+Tm + (Dr/Dwtrc)(1 + fcr)) ∗ Pi(15)

If the device is not connected with selected RSU, the RSUwill send a feedback to the cloud. If the mobile device sendsrequest to a RSU for the result, then the cloud will deliverthe result to the device through the current RSU. In this case,the round-trip delay is given as:

Tdel22 = Tdel21+Tf +Tr1+Tr2+(Dr/Dwtrc)(1+fcr) (16)

where Tf is the delay for sending feedback from the RSU tothe cloud, Tr1 is the delay for sending request by a mobiledevice for result to the RSU, under which the device ispresent, and Tr2 is the delay for forwarding the request bythe RSU to the cloud. The power consumption of the mobiledevice during this period is given as:

Pdel22 = Pdel21 + Tr1 ∗ Pa + (Tf + Tr2

+(Dr/Dwtrc)(1 + fcr)) ∗ Pi(17)

If ps and pu are the probabilities of the presence of themobile device under the predicted RSU, the round-tripdelay is given as:

Tdel2 = ps ∗ Tdel21 + pu ∗ Tdel22 (18)

The power consumption of the mobile device during thisperiod is given as:

Pdel2 = ps ∗ Pdel21 + pu ∗ Pdel22 (19)

However, though we have considered the case that themobile device may not be present under the coverage ofthe predicted RSU, the probability of this case is verylow, because the cloud is dynamically maintaining the usermobility information. As the user current location and thecurrent RSU serving the mobile device is predicted in oursystem, the delay in delivering the result is reduced. Ac-cordingly, the power consumption of the mobile device isreduced.

4.2 Delay model for Information Processing in CloudFrom the previous subsection the total delay in data trans-mission between RSU and mobile device (Tt) has been

11

determined using equation (10). Now, if cloud performsdata processing, then the uplink data transmission delaybetween RSU and cloud is given as:

Trc = (1 + frc) ∗ (Dc/Uptrc) (20)

The downlink data transmission delay between RSU andcloud is given as:

Tcr = (1 + fcr) ∗ (Dr/Dwtrc) (21)

Therefore, the total data transmission delay between mobiledevice and cloud is given as:

Ttn = Tt + Trc + Tcr (22)

The data processing delay inside the cloud is given as:

Tc = Dc/dpc (23)

The cloud after processing the data, predicts the current lo-cation of the user by analysing the mobility information andaccordingly the RSU currently serving the mobile device.After that the cloud sends the result to that RSU along withdevice ID and request ID. Hence, the round-trip delay is:

Tdel31 = Ttn + Tc + Tm (24)

where Tm is the delay in predicting the current location andthe RSU serving the device currently. The power consump-tion of the mobile device during this period is given as:

Pdel31 = Tt ∗ Pa + (Trc + Tcr + Tc + Tm) ∗ Pi (25)

If the device is not connected with the selected RSU, theRSU will send a feedback to the cloud. If the mobile devicesends request to a RSU for the result, then the cloud willdeliver the result to the device through the current RSU. Inthis case, the round-trip delay is given as:

Tdel32 = Tdel31+Tf +Tr1+Tr2+(Dr/Dwtrc)(1+fcr) (26)

where Tf is the delay for sending feedback from the RSU tothe cloud, Tr1 is the delay for sending request by a mobiledevice for result to the RSU, under which the device ispresent, and Tr2 is the delay for forwarding the request bythe RSU to the cloud. The power consumption of the mobiledevice during this period is given as:

Pdel32 = Pdel31 + Tr1 ∗ Pa + (Tf + Tr2

+(Dr/Dwtrc)(1 + fcr)) ∗ Pi(27)

If ps and pu are the probabilities of the presence of themobile device under the predicted RSU, the round-tripdelay is given as:

Tdel3 = ps ∗ Tdel31 + pu ∗ Tdel32 (28)

The power consumption of the mobile device during thisperiod is given as:

Pdel3 = ps ∗ Pdel31 + pu ∗ Pdel32 (29)

However, as the cloud is dynamically maintaining the usermobility information, the probability of the case that themobile device is not present under the coverage of the pre-dicted RSU is very low. As in the proposed model the RSUunder which the user is currently present is predicted, thedelay in delivering the result is faster and correspondinglythe power consumption of the mobile device is reduced.

5 PERFORMANCE EVALUATION

5.1 Mobility Dataset

The mobility dataset3 is collected from 100 mobile usersfrom their GPS-enabled smart phones and Google Map time-line for 6 months in the Kharagpur and Kolkata region ofIndia. The dataset consists of timeseries data of GPS traceswith the total time-duration of 26,8041 hours. The GPSpoints are logged in a high-sampling rate of 60-75secs.

5.2 Experimental Set-up

We aim to demonstrate the efficacy of Mobi-IoST with thereal-life mobility dataset. The performance analysis hasbeen carried out in following aspects: (i) movement patternmodelling, (ii) next location (and sequence) prediction and(iii) route prediction given the source and destination pair.Typically, accuracy, recall and F-measure are used to evaluateMobi-IoST and six baseline methods are implemented tocompare with our method. 70% of the movement traces areused for modelling, 20% and 10% for testing and validatingrespectively. The location sequence prediction task is evalu-ated in different time-scales, from 5mins to 60mins. The pathprediction task has been carried out in seven different time-bins (commuting time) (i) >=10mins, (ii) >10 and <=15mins, (iii) >15 and <=20 mins, (iv) >20 and <=30 mins,(v) >30 and <=40mins, (vi) >40 and <=45 mins and (vii)>45 and <=50mins. The intuition is to evaluate the efficacyof Mobi-IoST with respect to different trips with varyingcommuting time. For this purpose, the trips are dividedinto such seven classes based on their commuting time.The intuition is to evaluate the efficacy of Mobi-IoST withrespect to different trips with varying commuting time. Forthis purpose, the trips are divided into such classes basedon their commuting time. For each such classes, the tenfoldcross validation policy has been deployed where all tripswithin the same class are randomly divided into ten folds,where nine folds are utilized for training and one fold forvalidation. It guarantees that any trip in the validation setwill not appear in the training set. Next, the predictionaccuracy for all the trips in the validation set are computedand the average value of the accuracy measure for all sevenclasses are reported.

5.3 Movement Analysis

The performance measurement of the movement analysismodule have been carried out by three measurements,namely, accuracy, recall and F-measure. Apart from that,we evaluate the performance of the movement behaviourmodelling framework by comparing with six baseline meth-ods, semantic trajectory modelling [16], Bayesian network [29],Longest Common Sub-Sequence (LCSS) [27], Markov chain[30], Convolutional Neural Network Approach [18] and Spatio-temporal Recurrent Neural network (ST-RNN) [17]. It may benoted that the trajectory modelling modules of all of thecited works have been implemented with our dataset todepict the effectiveness of our framework.

3. Sample dataset available:https://drive.google.com/drive/folders/1BpM-K3clH6XYpSHkFe12aGsG8n1AclI4?usp=sharing

12

One of the major challenge of the proposed framework isto reduce the delay in delivering the processed information,and therefore if new GPS trace comes, the system shouldbe able to re-learn the pattern effectively. TABLE 2 showsthe performance measurements (accuracy, learning and re-learning time) compared to six baseline methods. The car-dinality (number of trajectory-windows × day) of the testdata of each agent for learning and re-learning are 18× 150and 6 × 20 respectively. Vlachos et al. [27] propose non-metric similarity function LCSS by computing the similaritybetween trajectory segments. Although the method outper-forms other distance metrics, such as, Euclidean or DTW,but it only calculates the topological or geometrical simi-larity ignoring the semantics of the trajectories. Semanticenrichment of trajectories and modelling to predict nextlocation has been studied in [16]. Mingqi et al. utilizesthe Bayesian network place classifier [29] to to categorize thesemantic of stay-points. Cheng et al. [30] model the check-insequences of individuals using matrix factorization methodand markov chain for personalized POI recommendations.Recently researchers are devoted to deploy neural networks[17], [18] to predict the next location sequences accurately.However, the neural network based methods are costly interms of re-learning and stability.

It is observed from TABLE 2 that our framework, Mobi-IoST outperforms all other baselines, except ST-RNN byapproximately 10-18%. Mobi-IoST not only provides nextlocation prediction based on some prediction technique(such as, CNN, RNN or Markov-model), rather it modelsindividual’s movement patterns over days, captures thefrequent path followed in several contexts and makes thenext location sequence prediction. Further, the learning andre-learning rate as well as stability of our framework issignificantly better compared to others. It is observed thatthe learning and re-learning rates of CNN and ST-RNN aresignificantly higher by 10-20min and 14-24min than Mobi-IoST respectively. These measurements are important forour case, since the system needs to incorporate any suddenmovement pattern change of user effectively. In summary,the deep learning architectures used in the existing worksare computationally extensive, and it is shown that suchdeep architecture may not be beneficial for time-criticalapplications, where a delay-aware solution is necessary.

Figs. 4-6 show the performance metrics of Mobi-IoSTwith the existing work [17]. The accuracy to predict stay-point information is represented in Fig. 6. It can be observedthat Mobi-IoST has gained 88% to 95% for trajectory stopsequences ranging from 2 to 10. There is a significantdrop of accuracy percentage from 93% to 80% of [17] inthe same set-up. The key reason behind this observationis the proposed movement modelling named User move-ment graph, where user movement pattern is modelled in amulti-layer graphical model. The frequent path network (FPN)(layer 4 of the User movement graph) learns the movementpaths frequently followed by the user deploying DynamicBayesian network. Thus, the model captures all sequencesof paths visited in different spatial and temporal contexts.Next, in the proposed model, k-order markov chain is usedto effectively model the spatio-temporal regularity of thetrajectory sequences, where the next location depends on krecent observations. On the other side, the existing work [17]

Fig. 4: Recall values for location prediction based on time-stamp value (min)

Fig. 5: F-measure values for location prediction based on time-stamp value (min)

Fig. 6: Accuracy percentage for prediction trajectory sequences

Fig. 7: Accuracy values for path prediction given source anddestination

use deep architecture to capture the spatial and temporalcontextual information, however fall short to predict longsequences of trajectory stay-point or stop points. Fig. 4and Fig. 5 show the recall and F-measure values of locationsequence prediction in different time-scale, from 5mins to

13

TABLE 2: Performance comparison of Movement modelling module of Mobi-IoST with baseline methods

Metric Semantic model [16] Bayesian [29] network LCSS [27] Markov predictor [30] CNN [18] ST-RNN [17] Mobi-IoSTAccuracy 84.02% 78.67% 72.95% 80.23% 86.08% 91.7% 93.25%

Learning (min) 8.4 10.2 16.8 10.6 32.4 22.8 12.6(|TrajW | : 18× 150)

Re-learning (min) 3.8 6.6 14.2 6.2 28.6 18.6 4.2(|TrajW | : 6× 20)

60mins. Since the major aim of this work is the efficient de-livery of service while the agent is on move, the experimen-tal evaluation based on the time-stamp value is justified. Wehave found the recall and F-measure values in the range of0.95 to 0.81 and 0.96 to 0.78 for twelve time-stamp valuesrespectively. The results indicate that Mobi-IoST not onlyprovides accurate predictions, but performs better than ST-RNN [17] while the time-stamp values increase as shown infigures 4 and 5. The reason behind the consistent predictionresult with increased time-stamp value is that we haveconsidered both spatial location and temporal span of avisit-sequence to model the proposed FPN of user movementgraph. The prediction algorithm is capable to predict nextlocation sequences based on both recent k locations andthe time-duration of each stay-points. Thus, the frameworkis suitable to predict longer sequences of locations moreeffectively than existing work. Fig. 7 represents the accuracyof predicting path given source and destination, where x-axis is the commuting time. The experimental set-up isdiscussed in section 5.2. All of the performance measurevalues are computed based on the correct number of grid-id predictions. The accuracy percentage lies in the range of89% to 95.3% for seven time-bins.

The performance evaluation is carried out to validatewhether Mobi-IoST is suitable to deliver the processed in-formation to the user-device based on user mobility predic-tion. To assist users and make intelligent decisions in time-critical applications, it is very crucial to model and predictusers’ next locations apriori based on varied spatio-temporalcontexts. There are several challenges such as (i) how ac-curately the model predicts next location sequences (notonly the immediate next location), (ii) whether the model iscapable to accommodate any sudden change of movementpattern of users. It has been observed that our frameworkhas outperformed in all of the performance measurementscompared to the baseline methods. The experimental resultspresent that to predict user’s next location Mobi-IoST takesapproximately 4.23-9.02sec, which is used in section 5.4 todetermine the delay and power consumption in Mobi-IoST.

5.4 Delay and Power Consumption

The mobile device sends data to the RSU for processing. TheRSU/cloud performs processing and sends back the resultto the mobile device. The mobile device may move to thecoverage of another RSU before getting the result. In suchcases, the connection interruption period is considered 10-30sec. MATLAB2015 is used for the simulation. The parametervalues considered in this analysis are presented in TABLE 3.In this analysis we consider the following two cases:

• Information processing inside the RSU• Information processing inside the cloud

TABLE 3: Simulation parameters

Parameter Valueu 40-85km/hr

Uptmr 50MbpsDwtmr 70MbpsUptrc 75MbpsDwtrc 150Mbpsfmr 0.008-0.25frm 0.008-0.25frc 0.05-0.5fcr 0.05-0.5dpr 500Mbpsdpc 1GbpsTm 4.23-9.02secPa 0.11WPi 0.055W

In the first case, we have considered health parameterdata transmitted by the mobile device. Blood pressure level(systolic and diastolic), body temperature, pulse rate ande.c.g. data are considered. The RSU after processing thehealth data, sends back the current health status (nor-mal/abnormal) as result to the mobile device. If abnormalityis detected, then the parameters which seem to be abnormalare also notified in the result. Here, a preliminary healthchecking is performed by the RSU to predict the healthstatus. The collected health parameter values are comparedto the normal range, e.g. the normal blood pressure range,normal body temperature, normal pulse rate etc. If eachof the collected value falls within the normal range withrespect to the current ambience and his/her health profile,then health status is predicted as normal. Otherwise, thehealth status is predicted as abnormal. The amount of datatransmission to serve each user request is considered 70-90 KB. The round-trip delay and power consumption ofthe mobile device (user-device) in the proposed approach,are presented in Fig.8 and Fig.9. The delay and power arecompared with the existing mobility-aware task delegationmethod [3]. In our approach the RSU works as a fog deviceand performs the data processing. If the device moves tothe coverage of another RSU, the cloud predicts the currentRSU based on user mobility information. In conventionalmethod, the cloud performs data processing and the userreceives the result through RSU. However, if the user getsdisconnected due to movement to another RSU, the userhas to access the cloud to retrieve the result by serializingsession information [3]. But in our approach, the cloud itselfsends the result to the RSU, that is currently serving thedevice. The RSU then forwards the result to the mobiledevice. Hence, the delay and power consumption of themobile device in the proposed system Mobi-IoST are lessthan the existing system [3]. This is observed that Mobi-IoST reduces the delay and power by approximately 23-26%and 37-41% respectively than the existing method [3].

In the second case, we have considered video data

14

Fig. 8: Round-trip delay in proposed and existing methods

Fig. 9: Power consumption of mobile device in proposed andexisting methods

Fig. 10: Round-trip delay in proposed and existing methods

Fig. 11: Power consumption of mobile device in proposed andexisting methods

transmitted by the mobile device. The RSU after process-ing the video data, sends back the processed data to themobile device. The amount of transmission to serve eachuser request is considered 2-20 MB. The round-trip delayand power consumption of mobile device in the proposedapproach, are presented in Fig.10 and Fig.11. The delay andpower are compared with the existing mobility-based taskdelegation method [3]. In our approach the cloud performsthe data processing. After that based on user geo-locationinformation, the cloud predicts the current location of theuser and the RSU currently serving the device. The cloudforwards the result to the RSU, which then sends back theresult to the mobile device. However, in the existing method,the cloud performs data processing and the user receives theresult through RSU after accessing the cloud. Moreover, ifthe user gets disconnected and moves to the coverage of an-other RSU, the user has to access the cloud through the newRSU to retrieve the result by serializing session information.Whereas in our approach, the cloud itself sends the resultto the RSU, that is currently serving the device. The RSUthen forwards the result to the mobile device. Hence, thedelay and power consumption of the mobile device in theproposed system are less than the existing system [3]. Thisis observed that the proposed system reduces the delay andpower by approximately 55-60% and 57-74% respectivelythan the existing system [3]. This is observed that for smallas well as large scale processing, our Mobi-IoST reduces thedelay and power consumption of the mobile device. As aresult, the QoS is enhanced.

6 CONCLUSIONS

Seamless connectivity is a major challenge during data andcomputation offloading in any mobile network. The process-ing of raw data collected using IoT devices and delivery ofthe result to the client mobile device becomes a challengeif the client frequently changes location. In this paper, wehave proposed a real-time cloud-fog-edge IoT collaborativeframework, namely Mobi-IoST, for efficiently delivering theprocessed information to the user-device based on user mo-bility prediction and intelligent decision making. The mobiledevice acts as an edge device, and the RSU is used as fogdevice for processing the raw data collected by the mobiledevice from the IoT devices. If the user changes location andgets disconnected from the RSU, it forwards the result tothe cloud. The cloud determines the current location of theuser based on the mobility information and delivers the re-sult accordingly. The mobility prediction module primarilystores the movement traces and models the frequent pathfollowed by the individual in different contexts. Further,it deploys a hidden markov model based location predictorfor efficiently predicting the location sequences. The real-life data of movement traces yield approximately 10-18%improvement compared to other existing methods. More-over, the simulation results demonstrate that the proposedfog computing framework reduces the delay and power byapproximately 23-26% and 37-41% respectively than the ex-isting mobility-aware task delegation system. In the future,the Mobi-IoST framework will be extended to capture thecellular data usage patterns from such time-series data andprediction of location sequences may help to appropriately

15

manage the power and bandwidth related resources. Theproposed Mobi-IoST will act as a foundation of mobility-aware network resource management in the future.

REFERENCES

[1] J. Gubbi, R. Buyya, S. Marusic, and M. Palaniswami, “Internetof things (iot): A vision, architectural elements, and future direc-tions,” Future Generation Computer Systems, vol. 29, no. 7, pp. 1645–1660, 2013.

[2] M. Chiang and T. Zhang, “Fog and iot: An overview of re-search opportunities,” IEEE Internet of Things Journal, vol. 3, no. 6,pp. 854–864, 2016.

[3] A. Mukherjee, D. G. Roy, and D. De, “Mobility-aware task delega-tion model in mobile cloud computing,” The Journal of Supercom-puting, vol. 75, no. 1, pp. 314–339, 2019.

[4] S. Ghosh and S. K. Ghosh, “Thump: Semantic analysis on trajec-tory traces to explore human movement pattern,” in Proceedingsof the 25th International Conference Companion on World Wide Web,pp. 35–36, 2016.

[5] C. Chen, D. Zhang, X. Ma, B. Guo, L. Wang, Y. Wang, andE. Sha, “Crowddeliver: planning city-wide package delivery pathsleveraging the crowd of taxis,” IEEE Transactions on IntelligentTransportation Systems, vol. 18, no. 6, pp. 1478–1496, 2017.

[6] K. A. Eldrandaly, M. Abdel-Basset, and L. A. Shawky, “Internet ofspatial things: A new reference model with insight analysis,” IEEEAccess, vol. 7, pp. 19653–19669, 2019.

[7] N. Fernando, S. W. Loke, and W. Rahayu, “Mobile cloud comput-ing: A survey,” Future Generation Computer Systems, vol. 29, no. 1,pp. 84–106, 2013.

[8] J. Zhang, X. Hu, Z. Ning, E. C.-H. Ngai, L. Zhou, J. Wei, J. Cheng,and B. Hu, “Energy-latency tradeoff for energy-aware offloadingin mobile edge computing networks,” IEEE Internet of ThingsJournal, vol. 5, no. 4, pp. 2633–2645, 2018.

[9] K. Gai, M. Qiu, H. Zhao, L. Tao, and Z. Zong, “Dynamic energy-aware cloudlet-based mobile cloud computing model for greencomputing,” Journal of Network and Computer Applications, vol. 59,pp. 46–54, 2016.

[10] Y. Zhu, Q. He, J. Liu, B. Li, and Y. Hu, “When crowd meetsbig video data: Cloud-edge collaborative transcoding for personallivecast,” IEEE Transactions on Network Science and Engineering,2018.

[11] Q. Fan and N. Ansari, “Towards workload balancing in fogcomputing empowered iot,” IEEE Transactions on Network Scienceand Engineering, 2018.

[12] C. Yang, Y. Liu, X. Chen, W. Zhong, and S. Xie, “Efficient mobility-aware task offloading for vehicular edge computing networks,”IEEE Access, 2019.

[13] C. Wang, Y. Li, and D. Jin, “Mobility-assisted opportunistic com-putation offloading,” IEEE Communications Letters, vol. 18, no. 10,pp. 1779–1782, 2014.

[14] F. Tang and H. Zhang, “Spatial task assignment based on informa-tion gain in crowdsourcing,” IEEE Transactions on Network Scienceand Engineering, 2019.

[15] H.-Q. Wu, L. Wang, and G. Xue, “Privacy-aware task allocationand data aggregation in fog-assisted spatial crowdsourcing,” IEEETransactions on Network Science and Engineering, 2019.

[16] J. J.-C. Ying, W.-C. Lee, T.-C. Weng, and V. S. Tseng, “Semantictrajectory mining for location prediction,” in Proceedings of the19th ACM SIGSPATIAL International Conference on Advances inGeographic Information Systems, pp. 34–43, ACM, 2011.

[17] Q. Liu, S. Wu, L. Wang, and T. Tan, “Predicting the next location: Arecurrent model with spatial and temporal contexts,” in ThirtiethAAAI Conference on Artificial Intelligence, 2016.

[18] A. Karatzoglou, N. Schnell, and M. Beigl, “A convolutional neuralnetwork approach for modeling semantic trajectories and predict-ing future locations,” in International Conference on Artificial NeuralNetworks, pp. 61–72, Springer, 2018.

[19] G. Pan, G. Qi, Z. Wu, D. Zhang, and S. Li, “Land-use classificationusing taxi gps traces,” IEEE Transactions on Intelligent TransportationSystems, vol. 14, no. 1, pp. 113–123, 2013.

[20] S. Ghosh and S. K. Ghosh, “Modeling of human movement behav-ioral knowledge from gps traces for categorizing mobile users,” inProceedings of the 26th International Conference on World Wide WebCompanion, pp. 51–58, 2017.

[21] B. Du, C. Liu, W. Zhou, Z. Hou, and H. Xiong, “Detectingpickpocket suspects from large-scale public transit records,” IEEETransactions on Knowledge and Data Engineering, vol. 31, no. 3,pp. 465–478, 2019.

[22] C. Song, Z. Qu, N. Blumm, and A.-L. Barabasi, “Limits of pre-dictability in human mobility,” Science, vol. 327, no. 5968, pp. 1018–1021, 2010.

[23] Y. Zheng, “Trajectory data mining: an overview,” ACM Transactionson Intelligent Systems and Technology (TIST), vol. 6, no. 3, p. 29, 2015.

[24] S. Shekhar and S. Chawla, Spatial databases: a tour, vol. 2003.prentice hall Upper Saddle River, NJ, 2003.

[25] C. E. White, D. Bernstein, and A. L. Kornhauser, “Some mapmatching algorithms for personal navigation assistants,” Trans-portation research part c: emerging technologies, vol. 8, no. 1-6, pp. 91–108, 2000.

[26] A. Mukherjee, S. Bhattacherjee, S. Pal, and D. De, “Femtocellbased green power consumption methods for mobile network,”Computer Networks, vol. 57, no. 1, pp. 162–178, 2013.

[27] M. Vlachos, D. Gunopoulos, and G. Kollios, “Discovering similarmultidimensional trajectories,” in 18th International Conference onData Engineering, p. 0673, IEEE, 2002.

[28] J. Yamato, J. Ohya, and K. Ishii, “Recognizing human action intime-sequential images using hidden markov model,” in Proceed-ings 1992 IEEE Computer Society Conference on Computer Vision andPattern Recognition, pp. 379–385, IEEE, 1992.

[29] M. Lv, L. Chen, and G. Chen, “Discovering personally seman-tic places from gps trajectories,” in Proceedings of the 21st ACMInternational Conference on Information and Knowledge Management,pp. 1552–1556, ACM, 2012.

[30] C. Cheng, H. Yang, M. R. Lyu, and I. King, “Where you like togo next: Successive point-of-interest recommendation,” in Twenty-Third International Joint Conference on Artificial Intelligence, 2013.

Shreya Ghosh received the B.Tech. degree fromIIEST Shibpur, India, in 2015. She is currently a Re-search Scholar with the Department of Computer Sci-ence and Engineering, IIT Kharagpur, India workingtowards her PhD. Her current research interests includespatial informatics, trajectory data mining and cloudcomputing. Shreya is the recipient of TCS fellowship.

Anwesha Mukherjee received M.Tech. and Ph.D.degrees from the Department of Computer Science andEngineering, West Bengal University of Technology, In-dia, in 2011 and 2018, respectively. She is currentlyworking as Research Associate in the computer sci-ence department of IIT Kharagpur. Her research areasincludes IoT, Fog computing, mobile network and mo-bile cloud computing. She has received Young ScientistAward from International Union of Radio Science in2014 at Beijing, China.

Soumya K Ghosh is currently a Professor with theDepartment of Computer Science and Engineering, IITKharagpur. He was with the Indian Space ResearchOrganization, Bengaluru, India. He He has authoredor coauthored more than 200 research papers in re-puted journals and conference proceedings. His currentresearch interests include spatial data science, spatialweb services, and cloud computing.

Rajkumar Buyya is a Redmond Barry DistinguishedProfessor and Director of the Cloud Computing andDistributed Systems (CLOUDS) Laboratory at the Uni-versity of Melbourne, Australia. He is also serving asthe founding CEO of Manjrasoft, a spin-off company ofthe University, commercializing its innovations in CloudComputing. He has authored over 625 publications andseven text books including ”Mastering Cloud Comput-ing” published by McGraw Hill, China Machine Press,and Morgan Kaufmann for Indian, Chinese and interna-tional markets respectively. He is one of the highly cited

authors in computer science and software engineering worldwide (h-index=118,g-index=245, 73,200+ citations). Dr. Buyya is recognized as a ”Web of ScienceHighly Cited Researcher” in 2016 and 2017 by Thomson Reuters, a Fellow ofIEEE, and Scopus Researcher of the Year 2017 with Excellence in InnovativeResearch Award by Elsevier for his outstanding contributions to Cloud computing.