Connected Vehicular Transportation: Data Analytics and Traffic-dependent Networking
Cailian Chen, Tom Hao Luan, Xinping Guan, Ning Lu, Yunshu Liu
11 Connected Vehicular Transportation: DataAnalytics and Traffic-dependent Networking
Cailian Chen, Shanghai Jiao Tong UniversityTom Hao Luan, Deakin UniversityXinping Guan, Shanghai Jiao Tong UniversityNing Lu, Thompson Rivers UniversityYunshu Liu, Shanghai Jiao Tong University
Abstract
With onboard operating systems becoming increasingly common in vehicles, the real-time broadband infotainmentand Intelligent Transportation System (ITS) service applications in fast-motion vehicles become ever demanding, whichare highly expected to significantly improve the efficiency and safety of our daily on-road lives. The emerging ITS andvehicular applications, e.g., trip planning, however, require substantial efforts on the real-time pervasive informationcollection and big data processing so as to provide quick decision making and feedbacks to the fast moving vehicles,which thus impose the significant challenges on the development of an efficient vehicular communication platform.In this article, we present TrasoNET, an integrated network framework to provide realtime intelligent transportationservices to connected vehicles by exploring the data analytics and networking techniques. TrasoNET is built upontwo key components. The first one guides vehicles to the appropriate access networks by exploring the information ofrealtime traffic status, specific user preferences, service applications and network conditions. The second componentmainly involves a distributed automatic access engine, which enables individual vehicles to make distributed accessdecisions based on access recommender, local observation and historic information. We showcase the application ofTrasoNET in a case study on real-time traffic sensing based on real traces of taxis.
I. I
NTRODUCTION
Our earth is facing the unstoppable increasing trend of vehicles. In United States, there are on average cars forevery , people. In China, the amount of vehicles is estimated to be million in 2020. The massive increase invehicles has brought a series of social and environmental issues to our cities and daily lives such as frequent trafficjams, vehicle crashes, throat-choking air pollution, etc. A sustainable, intelligent and green transportation system isthus of crucial importance. Towards this goal, one practical solution is to use the cutting-edge wireless informationand communication technologies to provide real-time transport-related information services to road administratorsand vehicles, namely Connected Vehicular Transportation System (CVTS) [1–3]. As a result, the transportationefficiency can be significantly improved with more smooth traffic flows and travellers can get informed for morewise route selections and enhanced travel experience. Furthermore, both Google and Apple released their mobileoperating systems for autos in 2014. It is estimated that the global Connected Car Market will reach . billion in April 27, 2017 DRAFT a r X i v : . [ c s . N I] A p r of all autos sold in 2016 will be connected. Therefore, it is foreseeable that in near future, connectedvehicles would embody a pragmatic solution towards Intelligent Transportation System (ITS).CVTS aims to make safer and more coordinated use of transportation networks. Such a system lies on the timelycollection of road traffic information, effective data analytics, and quick decision making and feedbacks to the trafficmanagement facilities and vehicles. Besides the roadside sensors (e.g. GPS, cameras, inductive loops, RFID andin-road reflectors) deployed in the city-wide for traditional traffic sensing, connected vehicles provide a new efficienttraffic monitoring method by the live data streams of large number of the off-the-shelf mobile terminals (e.g., on-board wireless communication facilities, smartphones, tachographs and wearable devices). It is envisioned that withthe adoptions of the embedded, tethered or smartphone integrated vehicular sensing and communication facilities,the demand of on board infotainment services would become more demanding, which eventually would generate alarge volume of data required for processing. Moreover, the large variety of data sources and applications requireCVTS with the assets of fast response and processing rate, and more importantly, with high accuracy, reliabilityand security.With the increasingly growing data in CVTS [4], there rise the fundamental engineering challenges from thefollowing three aspects: (i) big data collection from ubiquitous roadside and in-vehicle sensors in the city; (ii) deepdata analysis in traffic management center; and (iii) real-time decisions returned to traffic management facilitiesand vehicles. In this cycle, the first step is to timely, effectively and economically collect monitoring data fromthe ubiquitous sensors in the city. Compared to the dramatic improvement on technical tools for handling data [5],vehicular networking for CVTS applications has, however, adapted much slower towards the low-cost and efficientdata collection, which motives this article.In this article, we unfold our journey by first reviewing the impact of data analysis on some representative realtimetraffic-related services. The basic requirements for vehicular networking architecture are then identified. It is followedby a traffic-dependent network architecture for traffic data collecting and efficient service provisioning. Lastly, acase study is presented to show how realtime traffic estimation and timely network access can be implementedunder the proposed architecture. The main contributions of this paper are summarized as follows: • A novel Traffic-Social Network framework, called TrasoNET, is presented for the first time to build theconnection of realtime traffic and networking. Under this framework, the data analytics of CVTS take effectson the macroscopic, midscopic and microscopic network resource allocation and network access. Networkinginformation is an effective data source for traffic sensing. It makes data analytics and traffic-dependentnetworking mutually beneficial, which is the core idea of this work. • A new traffic-dependent network access scheme is developed with network access recommendation from higherlayer (network) and distributed automatic access decision-making in lower layer (terminal). It enables individualvehicles to make access decisions based on access recommender and local observation on network conditions. • A case study is presented to show the real data analytics for traffic estimation in Shanghai, China. Extensivesimulations demonstrate that TrasoNEt can effectively select optimum network to ensure QoS of vehicles/mobile devices, and network resource is fully utilized without network congestions by data offloading.
Fig. 1: The spacial-temporal average traffic density based on GPS dataset of taxies in Shanghai, China on Jan. 24,2013. II. D
ATA A NALYTICS FOR
CVTS A
PPLICATIONS
The continuous monitoring on movements (e.g., safety services by short range v2x communications), mobilityapplications and vehicles condition monitoring would result in exponential growth of diverse source data whichprovides a wealth of information that is valuable to traffic management parties, drivers, repair shops and automakers.In the following, we list some representative CVTS applications to describe their dependance on data analytics.
1) Real-time traffic estimation:
By using the moving vehicles or smartphones on-board vehicles for data sensing,and uploading the sensing reports (such as time, location and heading direction of vehicles) to the data center, therealtime traffic conditions of the roads, such as average running speed and traffic density, can be achieved by dataanalysis.The taxi/bus management system in Shanghai, China, represents a practical deployment of the CVTS platform.Around , taxis and buses of Shanghai are now equipped with the on-board GPS and sensors which periodicallyreport the vehicle information (GPS location, velocity, heading direction, passengers on/off) in cycles ranging from seconds to minutes. This results in million records transmitted to the traffic management center everydaythrough cellular networks, and enables multiple management purposes. Road traffic conditions can be estimatedefficiently by sparse sensing and advanced estimation methods. For example, compressive sensing and matrixcompletion based methods are reported in [6, 7] based on the GPS dataset of Shanghai.
2) Online navigation for connected vehicles:
Traffic prediction is more difficult than traffic estimation. Fortu-nately, through correlation analysis of big data, traffic patterns can be gleaned more easily, faster and clearly thanbefore. For example, a social proximity mobility pattern of vehicles is adopted in [8], i.e., each vehicle has arestricted mobility region around a specific social spot such as a financial and sport center. By using data analysis, the social spots of vehicles in the real-world can be identified. The traffic peak probably appear at the rush houraround the social spot (SP), which makes traffics predictable. Five SPs has been shown in Fig. 1. In anotherexample, the traffic can also be predicted from the message published in social networks, such as network groupevents (e.g., big show and football game) information including time, place, and number of attendees. The trafficcan be predicted to influx to the social places before the event and outflow after the event.The traffic also has strong correlations with online navigation services: the more people in a particular geographicplace search for the routes to a particular destination online, the more probably the traffic congestion happens onthe route to the destination. In this case, the traffic can be predicted with the data from search engine. The onlinenavigation server can provide more feasible path plans for vehicles.
3) Remote vehicle diagnostics and road condition warning:
Based on the data collected from cars, the driverscan arrange better service interval by taking their own driving habits, and predicted wear and tear into accountrather than conventional means based on defined number of kilometers. More importantly, valuable information onpotential vehicle or road hazards can be delivered to drivers in realtime. For example, if a number of vehicles’traction control systems are activated at the same time and place, the cars in this area can be warned about “icyconditions”, fuel-efficient driving in heavy traffic, and etc.
4) Fuel up or charge:
During travelling in the city, the electric vehicles/hybrid electric vehicles (EVs/HEVs)may make decision on what time and where to fuel up or load at location-specific charging piles by estimating thedriving mileage according to real-time traffic conditions. The information provided by CVTS is valuable to designefficient ways of resource management in smart grid system in the city [9].
5) Dynamic urban planning:
Sensory data from vehicles and mobile devices provide a pervasive way to under-stand how people use the city’s infrastructure and affect the city, including urban dynamics, energy consumption andenvironment impacts such as noise and pollution. The big data related to real-time traffic can be used to improvecity’s services. For example, the planning and management parties can estimate how residential and working areasin cities are connected temporally, what the dynamic correlation of traffic density and pollution level appears, andhow to reduce operational costs by optimizing planning. It also creates feedback loops with vehicles to reduceenergy consumption and environmental impact.III. F
EATURES OF D ATA A NALYTICS IN
CVTSFor the aforementioned applications, to explore the strong correlation among multi-source data is the key, whichhelps us capture the present road conditions and predict the future. The data collection for connected vehicles andrelated applications of CVTS distinguish with the traditional ones from the following three aspects: • From static sensing to dynamic sensing : As a large amount of traffic data can now be harnessed throughubiquitous roadside sensors, vehicles and mobile devices, static traffic sampling no longer makes as much sense.Moreover, due to the complicated traffic conditions, accurate traffic estimation and prediction can hardly beachieved on small random samples, and require as much data as possible. Connected vehicles make it possible. • From precise data to messy data [10]: In the applications, allowing for imprecision (for messiness) of datamay be an advantage, rather than a shortcoming. We can infer the vehicles’ direction, speed and position with messy GPS data thus traffic estimation can be improved to the level of predicting the traffic congestion in aparticular road rather than a region in the city by using 65 millions of “dirty” (or “noisy”) taxi GPS data ratherthan small precise samples from digital camera and loop detector [6]. The data could be messy if they coveras many streets as possible. • From parametric data to nonparametric data : Correlations are useful in a small-data world, but they really shinein the context of large volume data and/or big data. For traffic sensing and prediction, different (complicated)traffic models can be assumed to parameterize the relationship of traffic flows. However, multi-source data inCVTS allows us to pick a nonparametric model with simpler algorithms, and it results in more accurate thanthe sophisticated solution [6, 10].To summarize, for data analytics of connected vehicles, we can relax the standards of allowable errors andincrease messiness by combining different types of data information from different sources. In dealing with evenmore comprehensive datasets, we no longer need to worry so much about individual data points, but biasing theoverall analysis. Through them we can glean insights more easily, faster, and more clearly than before. Correlationsof multi-source data let us analyze the city traffic not by shedding light on its inner working but by identifying auseful proxy (e.g., the online navigation requests implies the possible congestion) for it. It is foreseeable that bigdata enhances the data analytics in CVTS to change the way of services provisioning in and enables connectedvehicles from multi-dimensions of traffic, vehicular network and ITS.IV. T
RAFFIC - DEPENDENT N ETWORKING FOR
CVTSTo engineer an efficient and economic network architecture is the foremost issue to facilitate the data collection,decision feedbacks and traffic related services.
A. Network Framework
The network framework is challenged by following issues: • From a traffic sensing perspective, even with the broad mobility of vehicles and the dense deployment of staticsensors on the road, it cannot be guaranteed that the traffic information for all the roads in all the time couldbe sensed. It therefore calls for an efficient and economical sampling way for traffic sensing. Crowdsensingby more vehicles could be one of the solutions in the very near future. • From a social perspective, the spatial distribution of traffic may follow a specific social pattern such as thepower-law distribution features for the traffic in Shanghai, China which is shown in Fig.1. The traffic densitydecays from the business hot spots towards the boarder of the vehicles’ mobility regions. One of the mainchallenges is finding a good guidance for wireless access to heterogeneous wireless networks (cellular networkand vehicular ad hoc networks) such that the distributed traffic-dependent service requests can be satisfied withgood Quality of Service (QoS)[11].In this article, we describe a Traffic-Social Network (TrasoNET) framework for CVTS as in Fig. 2 to supportthe crowdsensing and network access guidance according to realtime traffic and traffic pattern.
Fig. 2: Network architecture for connected vehiclesTrasoNET consists of three layers: access network layer , data aggregation layer and application layer . Inthe access network layer, sensory nodes, including vehicles and the mobile devices, could connect to roadsidecommunication infrastructures (e.g., cellular base stations and roadside units) and communicate through LTE/5Gcellular networks and/or vehicular ad hoc networks (VANETs). The static sensors (e.g. cameras, inductive loops,RFID and in-road reflectors) transmit data through wired communication. In the data aggregation layer, the roadsidecommunication infrastructure are connected to corresponding backbone routers. Data flows are combined throughthe so-called central controller sub-layer or so-called fog computing server, and further delivered to the cloudserver through Internet. In the application layer , the traffic management center (TMC) aggregates the collectedmulti-source data from cloud and analyze the data to estimate and predict the road traffic. The cloud also connectsto other service providers such that the traffic-related information can be fused out and provided in the applicationlayer. Different traffic-related services are then delivered to vehicles through cellular core network and regionalVANET.We elaborate on the four core components in the framework to highlight the characteristics in traffic-dependentnetworking. Infrastructure : The access infrastructures consists of the evolved NodeBs (eNBs) and RSUs. It is assumed eNBs cover the whole city, and the communication link between mobile device and eNB is more stable than that betweenmobile device with RSU. RSU is equipped with a wireless transceiver operating on DSRC and/or WiFi, and hencethe transmission range is small compared with eNB. But it provides high-rate transmission for mobile devices. Dueto the explosive growth of mobile data traffic, the cellular network nowadays is straining to meet the current mobiledata demand and faces an increasingly severe overload problem. RSU is not only an alternative for V2I (vehicularto infrastructure) communications, but also enables an offloading for cellular networks [8].
Mobile Devices : We do not discriminate what kind of mobile devices they are, but care about what networkthey access to. Normally, smartphones can connect to cellular network through LTE/5G and VANET infrastructuresthrough WiFi, while vehicles can additionally connect to VANET infrastructure and other vehicles through DSRC.Since WiFi and DSRC technologies can be applied to drive-thru connection when they are moving on the road[12], we propose an automatic network access engine in the mobile devices to offload data originally targeted forcellular networks, which is referred to as the automatic offloading engine.
Central Controller : The central controller is connected to base stations (e.g., eNBs for LTE), RSUs and Internetbackbones. It allocates the network radio resources based on the realtime traffic estimated by TMC, and servicedemands requested by mobile devices. It acts as an interface between the physical network routers and the networkoperators to specify network services. The controller builds a logical control plane separated from data plane.Different from Internet Protocol (IP) based networks, such a frame enables mobile devices to move between differentaccess interfaces without changing identities or violating specifications. The control function can be implementedby a protocol known as OpenFlow which enable controller to drive the access network edge hardware in order tocreate an easily programmable identity-based overlay on the traditional IP core.
Cloud : As the data analytics center for TMC and other service providers, the cloud receives data from statictraffic sensors and mobile devices, and analyzes them for traffic estimation and prediction. Other traffic-relatedservices are then analyzed based on the realtime traffic and data from other service providers. One key featureprovided by cloud is the access guidance for the mobile devices to facilitate the automatic offloading engine.As it is shown in Fig. 3, TrasoNET builds the connection of data collection, analytics and traffic-dependentnetworking from the following three aspects:1) Firstly, the traffic big data are collected from static and mobile sensors through access network of TrasoNET.The aforementioned static sensors (e.g., cameras and inductive loops) transmit the traffic data to RegionalTraffic Management Center through wired networks. Ubiquitous data from Mobile Devices (e.g., embedded,tethered or integrated on-board units, and smartphone) could be transmitted through wireless access network.For example, the probe vehicles (such as Taxis and buses) and floating cars (such as police cars from PublicSecurity Bureau) in the city could provide sparse GPS data for preliminary traffic estimation. Then the traffic-dependent networking mechanism to be introduced in Subsection IV-B could facilitate big data collectionfrom ubiquitous Mobile Devices. More data improves the traffic estimation and other traffic related services.2) Secondly, on the aggregation layer and application layer of TrasoNET, the data analytics provides the real-timeregional and global traffic conditions. It facilitates the Central Controller to allocate wide-area network radioresources (e.g., base stations, RSUs and Internet backbones) according to the estimated traffic density, speed,
Regional Traffic
Management Center A
Real-time traffic estimation
Data Analytics D a t a F u s i o n Static SensorsSCOTS
Camera
Mobile Sensors
Infrared/RadarBrake detectorGPSSpeedo meter
Accelerometer DistanceRunning statesLocationSpeedAccelerationOBU Traffic density
Remote vehicle diagnosticsFuel up or charge
Application
Data Collection T r a ff i c - d e p e nd e n t N e t w o r k i n g Regional Traffic Management Center BRegional Traffic Management Center N
Wired CommunicationWireless Access Network
Fig. 3: Data collection, analytics and applicationsacceleration and other information of vehicles/users in the city. Another key feature dependent on real-timetraffic condition is the regional network access guidance for Mobile Devices, which realizes locally networkresource management. As for Mobile Devices, the decision-making of network selection and handover can begiven locally by the guidance-based access mechanism for efficiency and offloading purposes. In this sense,the data analytics take effects on the macroscopic, midscopic and microscopic network resource allocationand network access.3) Thirdly, the various data from different network components provide complimentary data for deep dataanalytics. For example, the number of vehicles connected to an access point of wireless communicationcan reflect vehicle density, which can reduce the cost for satisfactory traffic estimation accuracy comparedto traditional sensing methods with digital cameras and loop detectors. Besides the realtime estimation, theTrasoNET facilitates online navigation, remote vehicle diagnostics, fuel up and charge and other emergingapplications.
B. Traffic-dependent Network Access Mechanism
The access control of networks is one of the key mechanism to guarantee the real-time CVTS applications. Inthe framework of TrasoNET, we give a guidance-based automatic access mechanism for efficient and offloadingpurposes. From the perspective of network access, the aforementioned four components map the phases of guidance,information push and distributed decision-making into Access Recommender Console, Broadcasting and Automaticoffloading Engine.
Access recommender console : To recommend an “optimum network” to vehicles based on multiple criteria, thecloud could apply intelligent computation methods to set the priority of network access for a specific region underrealtime traffic condition. The well-known Analytic Hierarchy Process (AHP) for multi-criteria decision [13] is one
Establish the Hierarchy Structure
Construct judgment matrix
Establish priorities among the elements of hierarchy
Pass the consistency check?
Calculate overall priorities for the hierarchy
Pass the consistency check?
Output
YesNo No OptimumNetwork
Bandwidth
Delay Cost
Traffic Density
Cellular Wi-Fi DSRCObject LevelElementsLevelAlternatives
Level
Yes
Fig. 4: Analytic Hierarchy Process for network selectionof good solutions. The logical flowchart of AHP algorithm is given in Fig. 4. The key steps are introduced in thefollowing. • Model the network recommendation problem as a hierarchy which contains the goal, alternatives for reachingthe goal, and criteria for evaluating alternatives. • Establish priorities among the elements of the hierarchy by making a series of judgments based on pair-wisecomparisons of these elements. The compared results construct a pair-wise comparison matrix A = a i j, i, j =1 , , · · · , n , where n is the number of criteria of second level,and every element a ij is based on a standardizedcomparison scale from equal importance to dominance . • The pair-wise comparison matrix should satisfy transitive preference and strength relations, it is necessary tocheck its consistency. Calculate consistency indicators
C.I. , random consistency indicators RI , and get theconsistency ratio CR = CI/RI . For example, consistency of judgement matrix is acceptable for the case of
CR < . . • Synthesize these priority vectors to construct an overall priority vector and check the consistency again.The final priorities of alternative networks for the “Optimum Network” can be got through the above algorithm.Traffic density is the critical factor and reflects the feature of vehicles’ mobility.
Distributed Automatic Access Engine : The engine operators in an automatic process shown in Fig. 5. The QoSrequirements ( (cid:104) data rate, delay, cost (cid:105) ) of various applications are registered with local observation of vehicle speedand the access recommender pushed through cellular network. The access option can be decided by analyzing theregistered information, the received signal strength (RSS) of communication links and the statistical knowledge in RegisterDecision MakingKnowledge Base Cellular AccessWiFi Access
DSRC Access
QoS Requirement Access Recommendation CellularV2I/V2RV2V/V2RAccess OptionsLocal ObservationFuzzy Inference
QoS
Achievable QoSCurrent QoS
Handoveror not ( ) f S
Low
High S DefuzzifierAchievable QoS ( ) f Q Q Low High1
Fuzzifier ( ) f A A DSRC ( ) f O O Wi-Fi 0Cellular
DSRC ( ) f R R Wi-Fi
Fig. 5: Distributed decision making processthe past. It is noted that the knowledge base is defined as
Q(cid:104)
Speed, Application, Access option , QoS | Access Recommender (cid:105) , which can be abbreviated as Q(cid:104)
S, A, O, Q | R (cid:105) . The knowledge base could be updated by the new achieved QoSperiodically. The trustworthiness on access recommender can be adapted according to local observation and achievedQoS (access trials or QoS in a specific accessed network) for device’s access decision-making (handover to anotheraccess network or not). The adapted process can be implemented by designing proper low-complexity algorithm inAPP such as in Fig. 6 through rule based inference [14] for decision-making. In this process, the aforementionedproximity traffic pattern, locations of infrastructure and RSS statistics are the preferences for consideration so thatthe rules could be logically given.In the following, fuzzy rules are powerful to represent the relation between the achieved QoS under accessednetwork and the criteria (cid:104) S,A,O,R (cid:105) for automatic access engine. In fuzzy theory, a rulebase is a function F that mapsan input vector into outputs. Here, the premise variables are set as the four factors (cid:104) S,A,O,R (cid:105) . The achievable QoSlevel is defined as the output. The membership function for each variable can be defined. It could be simplified intosingleton fuzzified levels for each premise variable. For example, set
Low and
High for S , classify V oice, T ext and
V ideo for A , and let Cellular, W iF i and
V AN ET for both of access recommender R and access option ofthe engine. An exemplary fuzzy rule with l levels of output could be as follows:Rule i : If S is Low , A is V oice , O is Cellular and R is Cellular , then the achievable QoS could be
Level l .Comparing the achievable QoS Level l through fuzzy decision-making and the achieved QoS Level c , we candecide whether or not to handover to the “optimum network”. Only if achievable QoS Level l is better in a certaindegree than the achieved QoS Level c , the handover happens.V. T RAFFIC SENSING AND TRAFFIC - DEPENDENT NETWORKING : A C
ASE S TUDY
In this section, we describe a prototype of TrasoNET based data analytics for realtime traffic sensing and serviceprovisioning. Based on the framework depicted in Fig. 2, the prototyping system consists in three basic components: probe vehicles (PVs), TMC, and a cloud server for traffic analysis and network access recommender. Fig. 6 showsthe system structure which is applied to estimate the traffic of Shanghai, China based on real GPS dataset of , taxies (as PVs) on Jan. 24, 2013. To offload the cellular data traffic, the city-wide WLANs have been developed inShanghai and the number of AP is over , including i-Shanghai free WiFi in important social spots. Thusthe cellular network and WLAN network form the access network layer. The TMC and cloud server are on thedata aggregation layer. The application considered in this case is the on-demand network service provisioning forvehicles in a certain region.Fig. 6: Networking scheme for urban traffic estimation in Shanghai, China (left); The designed Android APP forAutomatic Offloading Engine (right) A. Data analytics for traffic sensing in CVTS
The taxies in Shanghai generate sensing reports every seconds and report the readings of GPS, i.e. location,time of report, current speed and headings, to TMC through cellular networks. The TMC collects all the trafficreports and constructs a huge traffic matrix X = { x ij } , in which each entry x ij represents the traffic condition ofthe i -th road (e.g. average speed based on all the reports from PVs on the road) at the j -th duty cycle of a day.For example, the location of each report is matched to one road by map matching algorithm, data from differentPVs are fused to get the traffic matrix X . Since the PVs cannot cover all the roads for all the time, TMC needsto estimate the traffic of un-sampled road in the traffic matrix. Matrix completion is applied in [6, 7] with the lowrank property of the traffic matrix. The matrix completion based estimation could be computed in the cloud. Theestimation result is then sent to TMC for traffic management and message publishing to the vehicles in the citythrough the traffic bulletin board or information push through cellular network.The main idea of the real trace analytics is as follows. Firstly, estimate the values of average speed in the un-sampled roads [15] with the constraints of the temporal continuity and bound of the traffic data (the speed limit),respectively. Secondly, use the sampled data, together with the estimated data, to solve the optimization problemby minimizing the rank of the traffic matrix. The so-called HaTTEM algorithm is presented in [7]. E s t i m a t i on E rr o r ( % ) sr = 15%sr = 17.5%sr = 20%sr = 22.5%sr = 25%sr = 27.5%sr = 30%sr = 32.5%sr = 35%fitting curve Fig. 7: Estimation error VS average entropy under different sample rateHowever, the integrity analysis of sensing report about specific roads tells the fact that only of over , roads in Shanghai have sensing reports for only time of the day. There are not any GPS reports of taxiesor buses in 17% roads within a whole day [6]. The coverage of taxies’ traces in the city is quite uneven due tothe aforementioned social proximity. So we need floating cars (FCs, e.g. police cars from Public Security Bureauand patrol cars from Traffic Management Department) to provide more data. These cars don’t need to change thepatrolling area, but just adjust the patrolling path for better traffic sensing. It is well-known that the disorder ofsamples can be expressed by entropy. The relation between the entropy and the estimation error is seen in Fig. 7.By planning the paths of only controllable FCs for the whole city of Shanghai, China, even with the ofcurrent PV samples, the average entropy is reduced from . nats (unity of entropy) to . nats. Thus, it is seenfrom Fig. 7 that the estimation error could be reduced from to with the complimentary GPS data of FCs. B. Recommendation algorithm for Traffic-dependent networking
The PVs and FCs connect not only to TMC for management, but also frequently to Internet for providing moreemerging services, (e.g., the new free taxi calling services in Shanghai, China with mobile APPs called Diditaxiand Kuaidadi ). Access to Internet would become a standard feature of future motor vehicles. However, simplyusing the cellular infrastructure for vehicle Internet access may result in an increasingly severe data overloadingissue, which eventually would degrade the communication service performance of both traditional smartphone andvehicular mobile users. This advances of citywide free WLAN access in Shanghai make it possible to serve vehicularusers in the near future. This article provides an access network recommendation mechanism for different networkapplications based on the estimated traffic which could be achieved by the method in Subsection V-A.In order to demonstrates the feasibility of traffic-dependent networking, we provide in Fig. 8 the intelligentnetwork access system (INAS) for efficient and economical communication. INAS consists of network recommender (cid:41)(cid:85)(cid:84)(cid:90)(cid:75)(cid:94)(cid:90)(cid:3)(cid:56)(cid:75)(cid:86)(cid:85)(cid:89)(cid:79)(cid:90)(cid:85)(cid:88)(cid:95) (cid:52)(cid:75)(cid:90)(cid:93)(cid:85)(cid:88)(cid:81)(cid:3)(cid:56)(cid:75)(cid:86)(cid:85)(cid:89)(cid:79)(cid:90)(cid:85)(cid:88)(cid:95) (cid:56)(cid:75)(cid:87)(cid:91)(cid:79)(cid:88)(cid:75)(cid:83)(cid:75)(cid:84)(cid:90)(cid:3)(cid:56)(cid:75)(cid:86)(cid:85)(cid:89)(cid:79)(cid:90)(cid:85)(cid:88)(cid:95) (cid:41)(cid:75)(cid:84)(cid:90)(cid:88)(cid:71)(cid:82)(cid:3)(cid:41)(cid:85)(cid:84)(cid:90)(cid:88)(cid:85)(cid:82)(cid:82)(cid:75)(cid:88) (cid:39)(cid:73)(cid:73)(cid:75)(cid:89)(cid:89)(cid:3)(cid:56)(cid:75)(cid:73)(cid:85)(cid:83)(cid:83)(cid:75)(cid:84)(cid:74)(cid:75)(cid:88) (cid:14)(cid:39)(cid:46)(cid:54)(cid:15) (cid:47)(cid:84)(cid:76)(cid:85)(cid:88)(cid:83)(cid:71)(cid:90)(cid:79)(cid:85)(cid:84)(cid:3)(cid:45)(cid:71)(cid:90)(cid:78)(cid:75)(cid:88)(cid:79)(cid:84)(cid:77) (cid:57)(cid:75)(cid:88)(cid:92)(cid:79)(cid:73)(cid:75)(cid:3)(cid:56)(cid:75)(cid:86)(cid:85)(cid:89)(cid:79)(cid:90)(cid:85)(cid:88)(cid:95) (cid:51)(cid:85)(cid:72)(cid:79)(cid:82)(cid:75)(cid:3)(cid:39)(cid:54)(cid:54) (cid:39)(cid:73)(cid:73)(cid:75)(cid:89)(cid:89)(cid:3)(cid:43)(cid:94)(cid:75)(cid:73)(cid:91)(cid:90)(cid:79)(cid:85)(cid:84)(cid:39)(cid:91)(cid:90)(cid:85)(cid:83)(cid:71)(cid:90)(cid:79)(cid:73)(cid:3)(cid:39)(cid:73)(cid:73)(cid:75)(cid:89)(cid:89) (cid:11) (cid:76)(cid:91)(cid:96)(cid:96)(cid:95)(cid:3)(cid:82)(cid:85)(cid:77)(cid:79)(cid:73)(cid:3)(cid:17)(cid:3)(cid:82)(cid:75)(cid:71)(cid:88)(cid:84)(cid:79)(cid:84)(cid:77) (cid:12) (cid:52)(cid:75)(cid:90)(cid:93)(cid:85)(cid:88)(cid:81)(cid:3)(cid:3)(cid:3)(cid:3)(cid:3)(cid:3)(cid:3)(cid:3)(cid:3)(cid:3)(cid:57)(cid:75)(cid:82)(cid:75)(cid:73)(cid:90)(cid:79)(cid:85)(cid:84)(cid:58)(cid:75)(cid:88)(cid:83)(cid:79)(cid:84)(cid:71)(cid:82)(cid:3)(cid:56)(cid:75)(cid:86)(cid:85)(cid:89)(cid:79)(cid:90)(cid:85)(cid:88)(cid:95) Fig. 8: Block diagram of intelligent network access systemTABLE I: Comparison matrix based on AHP
Service Type Criterion Traffic Density Bandwidth Delay Payment Network PriorityTraffic Density 1 5 3 7 0.5558Bandwidth / / / / / / / / / / / / by TMC and automatic access engine in mobile devices. It works in three phases, i.e. information gathering, networkselection and access execution. The context repository module in Fig. 8 is the knowledge base in Fig. 5.Model the urban traffic as scalable grids. In the simulation, consider SPs and , vehicles in the areaof KM × KM with restricted mobility region for each vehicle. There are vertical and horizontal streets,respectively. Assume that vehicles mobility region is partitioned into multiple tiers co-centered at their SPs. Thedistribution of mobility follows social proximity model and the vehicle dense obeys the power-law decaying from thecenter of SP to the border of the mobility region with the exponent γ = 2 . Without loss of any generality, considertwo types of real-time applications, i.e. Voice Service and Video Service for individual vehicles. Assume the servicerequirements are minutes of voice and minutes of video on average. The data flow rate is . Kbps and Mbpsfor voice and video service, respectively. The data rates are RMB / Mb for cellular network and RMB / Gb permonth, respectively. Based on the aforementioned AHP method, the network access recommendation can be givenbased on the comparison matrices in Table I. It implies that Voice Service is sensitive to network access delay,while Video Service need more priority for bandwidth. Furthermore, we show the access network recommendationresult in Fig. 9 for different service types according to the traffic condition (vehicle density) demonstrated on thebottom X-Y layer of Fig. 9.For the traffic density, its second level pair-wise comparison matrix is formed as A × . The simulation result Vertical StreetsHorizontal Streets R ec o mm e nd a t i on I nd ex (a) Celluar Network for Voice Service Vertical StreetsHorizontal Streets R ec o mm e nd a t i on I nd ex (b) VANET for Voice Service Vertical StreetsHorizontal Streets R ec o mm e nd a t i on I nd ex (c) Celluar Network for Video Service Vertical StreetsHorizontal Streets R ec o mm e nd a t i on I nd ex (d) VANET for Voideo Service Fig. 9: Recommendation of cellular network and VANET for voice and video servicesabout the values of density-tolerance for the two applications (voice and video services) shows that without thenetwork selection algorithm, the successful transmission probability is nearly zero when traffic density is 0.04 forvoice service, and 0.06 for video service. The result shows that without the algorithm, it’s almost impossible tosatisfy every cars’ QoS need.It is noted that Fig. 9 represents the average index values for the recommendation of cellular network and VANET,respectively. If there are only two accessible networks, the priority can be normalized. It is easily seen from Fig.9 that cellular network is recommended for voice service in a large region around SP where the vehicular trafficdensity is relatively high. Therefore, cellular network is still the first choice for voice service, especially at SPs.On the other hand, VANET is recommended to offload the cellular network for video service in the region closeto SP (except SP due to QoS requirement). It indicates that the network selection/ handover is closely related tovehicular traffic condition, which is demonstrated the necessity of traffic-dependent networking.With the access network recommendation, the procedure of distributed automatic network access decision-makingcould be shown in Fig. 5.There are premise variables (cid:104) S, A, O, R (cid:105) , which represent S peed of vehicle, A pplicationof network (i.e. voice or vedio), current O ption of access network, and R ecommendation of access network,respectively. The output is achievable QoS represented by Q . The fuzzy sets and corresponding membershipfunctions for each premise variable can be seen in Fig. 5. S is in the range of ∼ km/h. The fuzzifier for A , O and R is singleton. Hence, we have the following fuzzy rules: • Rule 1: If S is Low , A is V oice , O is Cellular and R is Cellular , then Q could be level h ; • Rule 2: If S is Low , A is V oice , O is Cellular and R is V AN ET , then Q could be level h ; • · · · ; • Rule 16: If S is High , A is V ideo , O is V AN ET and R is V AN ET , then Q could be level l .With defuzzifier of fuzzy inference result, the distributed automatic network access engine determines the networkselection and handover. In order to avoid ping-pong handover due to the mobility and perturbation of QoS, set twothresholds for QoS and delay, respectively. Calculate the QoS improvement by switching the current network tothe other. Only if the improvement exceeds the QoS threshold for the time longer than the delay threshold, thehandover happens. VI. C ONCLUSION AND F UTURE R ESEARCH T OPICS
This article describes an architecture called TransoNET for data analytics and networking in connected vehiclesenabled transportation systems. To efficient manage network resources for CVTS applications, we describe thefeatures of data analytics and subsequently introduce the traffic-dependent networking approach for data collecting.It shows how vehicular traffic can be estimated by matrix completion and how the recommendation-automaticintegrated method provides efficient guidance to vehicles for network accessing. The data analysis based on realtraces of taxies gives an exemplary study on traffic sensing. In particular, we study a case of multiple-networkselection by the combination of network access recommendation from cloud and automatic access engine in vehicles.It has been demonstrated the necessity to explore the relationship between vehicular traffic and networking forproviding real-time services in CVTS.Based on the proposed CVTS architecture, potential research directions can be envisioned to improve the dataanalytics and networking performance from both cloud and vehicle sides. On the cloud side, big data processingalgorithm can be incorporated, e.g. crowdsourcing technologies, for ubiquitous traffic sensing such that morevehicles could take the roles of PV and FC. The social patterns of the vehicles may be considered to improvethe traffic crowdsensing. On the vehicle side, automatic network access engine needs low-complexity decision-making algorithms for explosively increasing infotainment services through vehicles to Internet connection. As theterminals of crowdsensing, the vehicles could be more intelligent by automatically adapting the cycles of sensingand reporting according to local vehicular traffic. We believe CVTS will attract enormous attention from academiaand industry in the near future. R
EFERENCES [1] M. Conti and S. Giordano, “Mobile ad hoc networking: milestones, challenges, and new research directions,”
IEEE Communications Magazine , vol. 52, no. 1, pp. 85–96, 2014.[2] K. Abboud and W. Zhuang, “Stochastic analysis of a single-hop communication link in vehicular ad hocnetworks,”
IEEE Trans. Intelligent Transportation Systems , vol. 15, no. 5, pp. 2297–2307, 2014. [3] H. T. Cheng, H. Shan, and W. Zhuang, “Infotainment and road safety service support in vehicular networking:From a communication perspective,” Mechanical Systems and Signal Processing , vol. 25, no. 6, pp. 2020–2038,2011.[4] J. Wan, D. Zhang, S. Zhao, L. T. Yang, and J. Lloret, “Context-aware vehicular cyber-physical systems withcloud support: architecture, challenges, and solutions,”
IEEE Communications Magazine , vol. 52, no. 8, pp.106–113, 2014.[5] M. Gramaglia, M. Calderon, and C. J. Bernardos, “Abeona monitored traffic: Vanet-assisted cooperative trafficcongestion forecasting,”
IEEE Vehicular Technology Magazine , vol. 9, no. 2, pp. 50–57, 2014.[6] R. Du, C. Chen, B. Yang, N. Lu, X. Guan, and X. Shen, “Effective urban traffic monitoring by vehicularsensor networks,”
IEEE Trans. Vehicluar Technology , vol. 64, no. 1, pp. 273–286, 2015.[7] R. Du, C. Chen, B. Yang, and X. Guan, “Vanet based traffic estimation: A matrix completion approach,” in
Proc. of 2013 IEEE Globecom , 2013, pp. 30–35.[8] N. Lu and X. Shen,
Capacity Analysis of Vehicular Communication Networks . Springer, 2014.[9] M. Wang, H. Liang, R. Zhang, R. Deng, and X. Shen, “Mobility-aware coordinated charging for electricvehicles in vanet-enhanced smart grid,”
IEEE J. on Selected Areas in Comm. , vol. 32, no. 7, pp. 1344–1360,2014.[10] V. Mayer-Sch¨onberger and K. Cukier,
Big Data: A Revolution That Will Transform How We Live, Work, andThink . Houghton Mifflin Harcourt, 2013.[11] M. D. Felice, R. Doost-Mohammady, K. R. Chowdhury, and L. Bononi, “Smart radios for smart vehicles:cognitive vehicular networks,”
IEEE Vehicular Technology Magazine , vol. 7, 2012.[12] T. H. Luan, X. Ling, and X. Shen, “Mac in motion: impact of mobility on the mac of drive-thru internet,”
IEEE Trans. Mobile Comput , vol. 11, no. 2, 2012.[13] T. L. Saaty and K. Peniwati,
Group Decision Making: Drawing out and Reconciling Differences . Pittsburgh,Pennsylvania: RWS Publications, 2008.[14] C. Chen, G. Feng, D. Sun, and X. Guan, “H-inf. output feedback control of discrete-time fuzzy systems withapplication to chaos control,”
IEEE Trans. Fuzzy Systems , vol. 13, no. 4, pp. 531–543, 2005.[15] C. Chen, S. Zhu, X. Guan, and X. Shen,
Wireless Sensor Networks: Distributed Consensus Estimation .Springer, 2014. B
IOGRAPHIES
Cailian Chen ([email protected]) is currently a Professor of Shanghai Jiao Tong University, China. Her research interests include vehicularad hoc networks, wireless sensor and actuator network and computational intelligence. Dr. Chen was one of the First Prize Winners of UniversityNatural Science Award from The Ministry of Education of China in 2007. She received the ”IEEE Transactions on Fuzzy Systems OutstandingPaper Award” in 2008. She was honored ”New Century Excellent Talents in University” by Ministry of Education of China, ”Pujiang Scholar”and ”Shanghai Rising-Star” by Science and Technology Commission of Shanghai Municipality, China.
Tom Hao Luan ([email protected]) recieved the B.Eng. degree from Xi’an Jiao Tong University, China, in 2004, M.Phil. degree fromHong Kong University of Science and Technology in 2007, and PhD degree from University of Waterloo, Canada, in 2012. He is currently aLecturer in the School of Information Technology at the Deakin University, Melbourne, Australia. From March 2013 to August 2013, he wasa visiting research scientist in the Institute of Information Engineering, Chinese Academy of Sciences. Xinping Guan ([email protected]) is currently a Distinguished Professor of Shanghai Jiao Tong University, China. He is also the Professorof ”Cheung Kong Scholar” Program, appointed by Ministry of Education of P. R. China, and the winner of ”National Outstanding YouthFoundation”, granted by NSF of China (NSFC). His current research interests include wireless sensor networks, cognitive radio and wirelesstechnologies for smart grid and smart community. He received First Prize Winners of University Natural Science Award from The Ministry ofEducation of China in 2006, and the Second Prize of National Natural Science Award from The Ministry of Science and Technology of Chinain 2008. He received the ”IEEE Transaction on Fuzzy Systems Outstanding Paper Award” in 2008.
Ning Lu ([email protected]) received the B.Sc. and M.Sc. degrees from Tongji University, Shanghai, China, in 2007 and 2010, respectively, and PhDdegree from University of Waterloo, Waterloo, Canada in 2015. He is currently an Assistant Professor in the Department of Computing Scienceat Thompson Rivers University, Canada. His research interests include capacity and delay analysis, media access control, and routing protocoldesign for vehicular networks.