WANG Jianqiang（王建強(qiáng)），LI Shiwei（李世威），ZHANG Yuzhao（張玉召）

School of Traffic and Transportation，Lanzhou Jiaotong University，Lanzhou 730070，China

Introduction

Internet of vehicles （IoV），the core application of Internet of things（IoT）in Intelligent transportation system（ITS），is leading in its technology and economic feasibility.IoV，referring to the system that the electronic tags embedded the vehicles via radio frequency identification technology，collects and makes effective use of the static and dynamic traffic information from vehicles，monitors the road conditions，and offers comprehensive services［1-3］.

Traditional road traffic data collection technology consists of fixed one and mobile one.The former includes magnetic induction，radar，and video recording while the latter is made up of active feedback technology（including GPS dynamic information and electronic tags）and passive collection technology （based on license plate automatic identification technology）［4］.The mode of data collection and exchanging in IoV differs from above-listed identification methods relying on induction coil or probe vehicles［5］.This study focuses on the identifying methods applied to urban traffic under the working environment of IoV based on its concrete analysis.

1 Working Environment of IoV

The embedded systems which monitor and control physical equipment via computing technology are internationally termed cyber-physical systems （CPS）or deeply embedded systems.The CPS node used in IoT can be divided into three types：passive CPS node，active CPS node，and Internet CPS node.In accordance to the specific demands on its used settings or services，active CPS node should be used as the major equipment mounted on vehicles and Internet CPS node as the major one for fixed facilities in IoV environment.Active CPS node is reputed to be excellent in storing，calculating，linking，and actively sensing abilities and equipped with directional or omni-directioinal antenna［6-7］.Above-mentioned functions apart，Internet CPS node is able to be linked with Internet，realizing its monitor and management via Internet，thus can be used in its fixed facilities in IoV.Based on CPS equipment，radio frequency identification might be adopted to read and exchange transportation information［8-9］.The information from every car can be read respectively，thus proliferates the amount of information［10］.Moreover，the contents acquired by conventional traffic flow detecting methods are only time series，which can't contain specific information of each car.On the contrary， IoV based on radio frequency indentification technology，is able to collect more detailed and accurate traffic flow information（see Fig.1）.

2 Analysis of Traffic Flow Parameters

Determine time interval T is a fundamental issue in identifying traffic conditions，which are largely influenced by factors like environment.The time intervals Tshould be set to be longer than the length of the maximum signal period，ranging from 5to 15min generally［11］.

The traffic flow parameters in IoV are collected and calculated differently from traditional ways，and the advanced technologies facilitate its accuracy［12］.In order to identify the traffic conditions exactly，three relevant parameters q（traffic flow），（mean specd of space），and k（density）are respectively used to have cluster analysis and judging traffic conditions.

Fig.1 IoV working environment

2.1 Traffic flow

Traffic flowq refers to the amount of vehicles passing one point of road within a period.Traffic flow of passing Internet CPS node nis counted by CPS node n.Supposing the traffic flow passing Internet CPS node nfromtato tb，q，is collected，the time interval Tshould meet Eq.（1）；the passing vehicle number within T，is the sum of all the vehiclespassing node n within fromtato tb.Among them，represents the vehicle numbered i entering in communicative zone of node nat time t and the value is 1.According to the definition of traffic flow，qis calculated as shown in Eq.（2）：

2.2 Mean speed of space

2.3 Density

Density（k）is the vehicle number of a road unit at an instant time and is calculated with corresponding Internet CPS node，with the unit of veh／km.Rtnis supposed to be the total sum of vehicles at time t within communicative zone of node n.The density is computed by Eq.（4）：

3 Identification Method of Traffic Conditions Using Fuzzy c-Means Based on Genetic Algorithm（GA-FCM）

3.1 Principle of fuzzy c-means（FCM）

FCM was initially proposed by Bezdek in 1981 for categorizing data points in multi-dimensional space.In fuzzy cluster analysis，that is a soft cluster method for traffic flow data，membership is used to signify the cluster extent of every data point［13］.

FCM divides nvectors xi（i＝1，2，…，n）into cfuzzy groups，and computes clustering center vj，minimizing objective function of non-similarity；the objective function is Eq.（5）：

where nis the number of the data points；cis the clustering groups number；Uis the membership matrix；V ＝｛v1，v2，…，vc｝is the clustering center matrix；uijis an element located in rowi and column j of matrix U，indicating the membership of data point i belonging to clustering center j，is the Euclidean distance between clustering center j and data point i；mis a weighed coefficient to control algorithm flexibility，m ∈（1，∞）.Assuming m ＝1，F(xiàn)CM is degenerated to hard c-means cluster，ranging from 1.5to 2.5for traffic flow data soft clustering and ideal value 2［14］.

If above mentioned conditions are met，Lagrange multiplier is constructed，then the partial derivative is calculated for all the input parameters，the essential conditions minimizing objective function values are shown in Eqs.（6）and（7）：

3.2 Algorithm improvement

The clustering results of FCM are largely influenced by the initial value，sensitive to outliers'data，and easily stuck in local extremum or saddle point，losing globally optimal solution，especially for the large sample cluster analysis.Genetic algorithm（GA）adopts the competing mechanism to seek optimal individual by genetic approaches such as selection，crossing，and differentiation.GA and FCM are combined to form solving algorithm GA-FCM，which sums the approximate global optimal solution.Then the approximate global optimal solution is used as the initial value of FCM and sums the exact value.

Firstly，vjis a gene，then V ＝（g1，g2，…，gc）is a chromosome，representing a sort of clustering results.gjis the real codes of clustering center j，able to shorten chromosome，which is useful for global exploring and convergence speed， yet complicated for the genetic operation.

The fitness function is shown in Eq.（8）：

where wis a constant and J（U，V）is displayed in Eq.（5）.Arithmetic crossover operator is used to cross，which means two parent individuals generate two sub-divisions.Algorithm randomly selects a record from data collections as a gene（a clustering center），and each chromosome is made up of randomc genes.s chromosomes are selected randomly to be the initial population as a minor part of samples collections，with its sizes differed due to the sizes of the data collections.Then individual is selected in proportion，with the basis of the chromosome fitness value and the current best chromosome is reserved.

3.3 Algorithm iterative process

The basic procedure of applying GA-FCM is listed as below.

Step 1：standardize the original data with range transform so as to unify its range and dimension.It is given by Eq.（9）：

Step 2：all the parameter initial values are set，c is the given clustering number，εis set as the threshold of iteration stopping，and gis set as the counter of iteration.

Step 3：real-encode the chromosome for clustering center in real numbers.

Step 4：subordinating degree function and fitness function are calculated to judge whether the given precision is reached and thus find out the best chromosome.

Step 5：selecting，crossing，differentiating，and judging whether genetic iterative times g is up to the bound or whether the best genetic clusters are found.

Step 6：the best chromosome with GA is the initial clustering center matrix V（0），Eq.（6）is used to calculate the membership matrix U（g），and Eq.（7）is to update the clustering matrix V（g＋1）.If ‖V（g＋1）-V（g）‖ ＜ε，or objective function J is less than a fixed threshold，then the algorithm is stopped，outputting the results；or clustering iterative process is proceeded until the end.

4 Simulation

4.1 Simulation scenario

Matlab was adopted to imitate the data processing way in IoV（see Fig.2）.

Fig.2 Simulation scenario

To classify the traffic flow situations，the data used in the emulation process should cover overall data varieties.

Some research conducted a characteristic analysis on the traffic flow data collected from relevant roads in Beijing，and elaborated on the statistical characteristics of urban road traffic flow and discussed on the emulation process［15］.

Urban traffic conditions are divided into four levels：non-congestion，slightly congestion，congestion，and heavily congestion levels.The time intervals were set to 5 min and data within a continuous week were collected，totally with 2 016 groups of data （see Fig.3）.During the emulation process，the value of crossover probability was 0.35，the mutation probability 0.05，the initial population s 200，m 2，andε10-6.

Fig.3 Traffic flow data

4.2 Simulation analysis

The results are showed in Fig.4and analysis of clustered traffic flow and speed is showed in Fig.5.The iterative curves of objective function are showed in Fig.6.

Fig.4 Results of cluster analysis

Fig.5 Results of clustering analysis

Fig.6 Iterative curves of objective function

The clustering centers of above-mentioned noncongestion，slightly congestion，congestion，and heavily congestion levels are［96.298 9 57.463 5 22.106 1］，［257.711 4 40.890 9 32.861 0］，［520.751 1 30.232 2 41.069 9］，and ［411.043 1 20.557 4 59.666 5］，respectively.Three parameters of the clustering centers denote traffic flow ［veh／（5min）］，speed （km／h），and density（veh／km）.The convergence rate of GA-FCM is faster than that of FCM，and iterative number of GA-FCM is less than that of FCM.Compared with FCM，GA-FCM is a more effective approach to realize the clustering in order to identify urban road traffic conditions precisely.

Premised on the condition that the accurate traffic flow data，the road conditions can be identified in real time.Speed largely influences the clustering results in fuzzy analysis，which accords with the traveling experience.High speed traffic speed is reputed as the non-congestion level，and the slow one the congestion level.Meanwhile，density plays an important role of differentiation，according with the traffic flow basic parameter model.Excess density value to the threshold is judged as the congestion level.Flow alone cannot be used to identify traffic conditions.It should be emphasized that the identifying traffic conditions should be assisted by three parameters together instead of only one of them.Under the condition of large quantities of traffic flow data，GA is helpful to improve the clustering accuracy in initial analysis and convergence speed.In some experiments，the traffic flow data was made cluster analysis without standardization.Although the cluster results were revealed，the sorting results，proved through the experiments，were inaccurate and largely influenced by several traffic flow parameters.

5 Conclusions

The present study explores the accurate identification method of road traffic conditions under the context of IoV.Based on the ultra-new communication technologies of IoV，real-time traffic flow data were collected，stored，and processed.Then，we can work out the basic traffic flow parameters for identification of traffic conditions.GA-FCM is adopted to sort the traffic flow data，prevented from traditional experience-based evaluation.Computer simulation is used to test its feasibility and effectiveness.

In face of new traffic communication environment，the ways，the traffic flow parameters are collected and calculated，are varied.So are the sorting methods for traffic conditions.Therefore， how to select more practical collecting and calculating methods，how to select sorting arithmetic as well as its scientific setting of its parameters still need further exploration.

［1］Wang J Q，Li S W，Zeng J W.Analysis of the Development Model of Internet of Vehicles［J］.Computer Technology and Development，2011，21（12）：235-238.（in Chinese）

［2］Liu X Y，Wu M Y.Vehicular CPS：an Application of IoT in Vehicular Networks［J］.Journal of Computer Applications，2012，32（4）：900-904.（in Chinese）

［3］Wang J Q， Wu C W.VANETs Opportunistic Routing Algorithm Based on Triangular Fuzzy Number ［J］.Application Research of Computers，2011，28（3）：1084-1087.（in Chinese）

［4］Wang J Q，Wu C W，Li X J.Research on Architecture and Key Technologies of Internet of Vehicles［J］.Microcomputer Information，2011，27（4）：156-158.（in Chinese）

［5］Wang L，F(xiàn)an Y Z，Zhang H.A New Traffic State Evaluation Approach Based on Probe Vehical and Multi-classification Fuzzy Pattern Recognition ［J］.Journal of Highway and Transportation Research and Development，2007，24（9）：92-95.（in Chinese）

［6］Sha L，Gopalakr I S，Liu X.Cyber2Physical Systems：a New Frontier［C］.2008IEEE International Conference on Sensor Networks， Ubiquitous and Trustworthy Computing，Taichung，Taiwan，China，2008：1-9.

［7］Easwaran A，Lee I.Compositional Schedulability Analysis for Cyber-Physical Systems［J］.ACM Sigbed Review，2008，5（1）：11-12.

［8］Ali K， Hassanein H.Passive RFID for Intelligent Transportation Systems ［C］.The 6th IEEE Consumer Communications and Networking Conference，Las Vegas，Nevada，USA，2009：1-2.

［9］Roberts C M.Radio Frequency Identification（RFID）［J］.Computers ＆Security，2006，25（1）：18-26.

［10］He Z C，Li Z T，Zhao J M.Traffic Flow Time Series Periodicity Based on Recurrence Quantitative Analysis［J］.Journal of Southwest Jiaotong University，2010，45（6）：946-951.（in Chinese）

［11］Yu J.On the Fuzziness Index of the FCM Algorithm ［J］.Chinese Journal of Computers，2003，26（8）：968-973.（in Chinese）

［12］Wang C L，Zhang C.Detection Model of Traffic Flow Parameters Oriented Internet of Vehicles ［J］.Computer Engineering and Applications，2012，48（23）：212-218.（in Chinese）

［13］Yang Z Y，Huang X Y，Du C H，et al.Study of Urban Traffic Congestion Judgment Based on FFCM Clustering［J］.Application Research of Computers，2008，25（9）：2768-2770.（in Chinese）

［14］Pal N R，Bezdek J C.On Cluster Validity for the Fuzzy c-Mean Model［J］.IEEE Transactions on Fuzzy System，1995，3（3）：370-379.

［15］Gao Z Y，Long J C，Li X G.Congestion Propagation Law and Dissipation Control Strategies for Urban Traffic ［J］.Journal of University of Shanghai for Science and Technology，2011，33（6）：701-708.（in Chinese）