Journal of Network and Computer Applications 36 (2013) 1308–1315

Contents lists available at ScienceDirect

Journal of Network and Computer Applications

1084-80

http://d

n Corr

E-m

journal homepage: www.elsevier.com/locate/jnca

CATS: An adaptive traffic signal system based oncar-to-car communication

Nitin Maslekar a,n, Joseph Mouzna a, Mounir Boussedjra b, Houda Labiod c

a Department of Instrumentation and Information Systems, irseem – Esigelec, Saint Etienne du Rouvray, Franceb VisionIT Group, Paris, Francec Department of Computing and Networks, ENST-Telecom ParisTech, Paris, France

a r t i c l e i n f o

Article history:

Received 14 July 2011

Received in revised form

28 December 2011

Accepted 29 May 2012Available online 15 June 2012

Keywords:

Adaptive traffic signal control

Clustering

VANET

45/$ - see front matter & 2012 Elsevier Ltd. A

x.doi.org/10.1016/j.jnca.2012.05.011

esponding author.

ail addresses: m.nitin@esigelec.fr, nitinmaslek

a b s t r a c t

Traffic signal controls play an important role in regulating vehicular flow at road intersections.

Traditional systems are not capable of adjusting the timing pattern in accordance with vehicular

demand. This results in excessive delays for road users. Hence it is necessary to develop dynamic

systems that can adjust the timing patterns according to traffic demand. In this paper, the design and

implementation of an adaptive traffic signal control system based on car-to-car communication is

presented. Also, a clustering algorithm is defined which will assist in estimating the density of vehicles

approaching an intersection. The cycle time, which is calculated using the estimated density of

vehicular traffic, helps in reducing both the waiting time for vehicles at intersections and queue length.

It is also shown that the proposed solution is collision free at intersections. The proposed system is

compared with a classic pre-timed system and an adaptive fuzzy logic system. The simulations also

show that the data convergence time and the communication delay between vehicles and traffic signals

do not compromise the efficiency of the system.

& 2012 Elsevier Ltd. All rights reserved.

1. Introduction

Recent years have witnessed an exponential increase invehicular traffic in urban scenarios. This has resulted in inefficienttraffic flow. Improper traffic signal control systems at roadintersections are major contributors to this inefficiency. In orderto ensure efficient traffic flow, it is essential to optimize trafficsignal control in accordance with traffic demand. Such optimiza-tions will not only result in smoother traffic flows, but also reducethe number of vehicles that stop at intersection. This will help toreduce travel time for users and also to cut vehicle emissions.

At present, most existing traffic control systems are static innature. These systems have a set pre-timed timing plans whichare generated offline using the statistical data of the trafficdemand. In certain cases the implemented timing plan may notbe optimal for the existing traffic conditions, thus limiting thecapabilities of offline optimization models for varying trafficdemand. Hence, it is necessary to develop an adaptive trafficcontrol system which will respond to vehicular demand andoptimize its timings accordingly. Such a system helps to easecongestion and its downstream effects without the cost andenvironmental impact of road expansion. In addition to this, it

ll rights reserved.

ar@ieee.org (N. Maslekar).

can also respond to unexpected or unplanned events, such as roadaccidents.

The flexibility provided by adaptive systems has propelledthem into a potential technique for improving traffic conditions inurban areas. Within this context there has been a great deal ofresearch carried out on these systems. Some of the adaptiveapproaches are mentioned in (Liu, 2007). But, to facilitate theadaptive procedure, they either involve installation of road sideinfrastructure or have high computational complexity. Conse-quently, there is a need for a cost effective and efficient methodfor adaptive traffic signal control. One approach that can help inthe implementation of adaptive solutions at a lower cost withlesser complexity is Vehicular Ad hoc Networks (VANETs). Thisapproach has gained a lot of interest among industrial andacademic research communities.

This paper presents an adaptive traffic signal control systembased on car-to-car communication (VANETs). Also, the paperdiscusses a clustering algorithm that is used to estimate thenumber of vehicles approaching an intersection. Based on theestimated density, an adaptive cycle time and the green time forthe different phases at the intersection are computed. The maingoal of this adaptive system is to reduce the average waiting timeexperienced by vehicles at intersections along with the reductionin the number of vehicles that stop at intersections.

The rest of the paper is organized as follows. Section 2 gives abrief overview of the existing adaptive traffic signal controlmechanisms. Section 3 presents the clustering algorithm used

N. Maslekar et al. / Journal of Network and Computer Applications 36 (2013) 1308–1315 1309

for density estimation. The design and implementation of theproposed adaptive system is discussed in Section 4. Section 5describes the simulations, results and analysis of the adaptivesystem. Finally, the conclusion and the future direction of thework are discussed in Section 6.

2. Related work

Traffic signal systems can broadly be classified as either staticor dynamic systems. Static systems are comprised of traditionalpre-timed systems. At pre-timed traffic signals each signal phaseis serviced in a programmed sequence that is repeated through-out the day. This kind of system can provide efficient operationwith the assumption that the signal timings reflect the currenttraffic conditions. However, vehicular flow is unpredictable andvaries in time. Under such situations the static cycle timings donot adjust to the variations in the flow. This creates traffic jams atroad intersections and brings about the need to develop dynamictraffic signal systems.

Of the various existing dynamic systems, adaptive controlsystems are currently the most advanced and complex. Theseare similar to traffic signals based on a responsive technique (Liu,2007) in which real-time data is used. But, instead of matchingcurrent conditions to an existing timing plan, the system uses anonline computer to create an optimal plan. Hence, the need for adatabase of the timing plans is eliminated. Amongst the numer-ous existing adaptive systems a distinction can be made accord-ing to the mode they use to determine the timing plan.

In (Wiering et al., 2004) an adaptive control system based onlearning algorithms is proposed. In this system the authorsdevelop a learning method with a road-user based function todetermine optimal decisions for each traffic light. This decision isbased on a cumulative vote of all vehicles waiting at an intersec-tion where each car votes using its estimated gain of settingthe light green. This way of co-learning allows the driver tochoose the route with the lowest expected waiting time. Thoughthis algorithm generally helps to reduce the waiting time atintersections, it may fail at isolated intersections. Also, this imple-mentation suffers from saturation because at a given moment alldrivers may choose the same optimal route and make the routecrowded. In a similar way, the authors in (Gregoire et al., 2007)describe an adaptive traffic light controller using intelligentagents at isolated intersections. The idea here is to let the agentslearn a traffic control policy by using machine learning algorithmsbased on the attribution of rewards. This algorithm has a highcomputational cost because to implement optimal timing pat-terns, the agents will need to analyze many subtle details of thesystem and this leads to a slow convergence of the optimumtiming pattern. Another mechanism of adaptive traffic signalcontrol based on the constraint satisfaction problem (CSP) methodis mentioned in (Mizuno et al., 2008). Here the signal parameters,the cycle length, the split and the offset are defined as CSP and theoptimal cycle length is determined by solving CSP with reductionin waiting time as the constraint. Though the simulation results ofshow that traffic jams are considerably reduced, solving the CSPrequires time to arrive at an optimal solution. In (Kutil et al., 2006)an adaptive algorithm is defined where queuing theory is used.Here an extended queue is modeled as a vector denoting thevehicles that have been waiting at an intersection. Including themean waiting times in the model allows for a fair traffic control.This helps to reduce the computation costs however it still takestime to come to an optimal solution.

To reduce the convergence time of obtaining an optimalsolution, (Amin and Jalili, 2006) defines a greedy control with ascarcity measurement approach (GSCM). In GSCM the control

method is similar to many feedback algorithms. This feedback isquantized to a value of resource scarcity. Since it adapts a greedyapproach, it may not reach an optimal solution if too manyfluctuations exist in the traffic network. One method based onfluid-dynamics and a car following model is defined in (Lammerand Helbing, 2008). This method defines a combination ofstabilization and optimizing rules by anticipating the arrival timeof vehicle platoons. These rules allow for a varying sequence oftraffic phases and a spatially coordinated operation to predict thevehicle flow. Such rule based approach helps in achieving anefficient traffic signal control.

For faster convergence to optimal timings, in (Huang andMiller, 2004) the concept of a smart intersection is introducedwhich makes use of wireless communication. In this system eachcar sends its information, such as speed and distance from theintersection, to the traffic signal. Using this information, trafficsignal control decides whether a car can go through the intersec-tion or not. Though this method decentralizes the control forfaster convergence, it is subject to failure of data communicationwhich may cause the whole system to collapse. But this workprovides the impetus to explore the domain of wireless technol-ogies for adaptive traffic control systems. With the advent ofwireless vehicular networks, these are looked upon as an effectivesolution for managing traffic at intersections. Of the few proposedsolutions, (Gradinescu et al., 2007) is one such solution where onehop car-to-car communications is used to implement trafficcontrols. In this contribution, traffic signal control listen to thecommunication between cars and estimate the density of vehiclesaround them and adjust the signal timings accordingly. In(Dresner and Stone, 2004), the authors describe a reservation-based intersection control mechanism where vehicles commu-nicate with the intersection and reserve a time slot for theirpassage. These works provide a base to further investigateVANETs for the implementation of adaptive traffic systems. Thismotivates the design and implementation of VANET-based adap-tive traffic signal control that is advocated in this paper, whichcan considerably improve traffic conditions by reducing thewaiting time at intersections when compared with static andother dynamic systems.

3. Density estimation through clustering algorithm

To implement an adaptive traffic signal system it is necessaryto gather information about the density of vehicles approachingan intersection. To accomplish this through VANETs, it is essentialto have an effective data dissemination strategy. In the pro-posed approach, a cluster-based data dissemination protocolcalled Clustering based on Direction In Vehicular Environment(C-DRIVE) (Maslekar et al., 2011) is designed and developed. Thisalgorithm is developed over the WAVE Short Message (WSM)protocol stack (Eichler, 2007). Before discussing the implementa-tion details of adaptive traffic lights, this section briefly describesthe concept and functioning of the C-DRIVE algorithm.

3.1. Assumptions

The clustering algorithm assumes that each participatingvehicle knows its own position using the Global PositioningSystem (GPS). Moreover, it considers that each vehicle isequipped with digital maps which enable them to determinethe direction of travel. Therefore, the direction information at theintersection can be computed a priori. It also assumes that eachvehicle is equipped with at least one wireless transceiver.

Fig. 1. Computing the travel direction at the intersection.

Fig. 2. Reference Points on the lane.

Fig. 3. Packet transmitted by Header.

N. Maslekar et al. / Journal of Network and Computer Applications 36 (2013) 1308–13151310

3.2. Cluster formation

In the proposed clustering algorithm, the formation is initiatedbased on the direction which the vehicle will take after crossingthe intersection. Fig. 1 presents a screenshot of a lane which issplit into different regions (regions between the dark lines inFig. 1). The size of these regions is equivalent to the radio range ofthe vehicles. With the GPS information, each vehicle computes itstravel path at the intersection based on the source and thedestination. More precisely, the direction it will take after cross-ing the intersection will be known.

For example, at the intersection (Fig. 1) a vehicle can take anyof the following directions (a) straight (S), (b) right (R) or (c) left(L). Using this information, clusters S, R and L are formed on aparticular lane. The cluster formation is initiated at a distance oftwice the radio range from the intersection and is identified bythe demarcation of the regions (shown as dark lines in Fig. 1).Once the cluster is formed, it is the responsibility of the electedclusterhead to compute and transmit the density information tothe traffic signal control.

To form clusters with minimum overhead and to elect theclusterhead, two approaches are adopted and compared. In thefirst approach, C-DRIVE defines an opportunistic direction basedpropagation function (Maslekar et al., 2009). In this approach, atthe initiating point, a vehicle sends a hello message to verify theexistence of any cluster for a particular direction in which it istravelling. If it receives a reply from a clusterhead, it joins thecluster. Otherwise it assumes itself as the first vehicle in thatparticular direction and acts as clusterhead for that cluster.Besides, to achieve accuracy in density estimation, a clusterheadmaintenance policy is defined to take into account the fact thatvehicles may overtake each other. In this approach, the mobilityof the vehicles within the cluster affects the stability andincreases the number of clusterheads in the network that isresulted from clusterhead changeovers.

In order to overcome this shortcoming in the second approach,a new clusterhead election policy to form stable clusters amongstthe vehicles is implemented. This policy reduces the number ofclusterheads in the network and the subsequent overhead due toclusterhead changeover. In this approach, referred to as ModifiedC-DRIVE (MC-DRIVE), a set of imaginary points termed Reference

Points (Fig. 2) are employed during cluster formation and cluster-head election. The cluster formation begins when a vehiclereaches the first reference point called Start Point and detectsthat it does not belong to any cluster. This indicates that it is thefirst vehicle travelling in a particular direction. This vehicleassumes the role of temporary clusterhead and is termed Header.This Header is responsible for the election of a clusterhead. It firstcomputes a threshold value: THDISTANCE and generates a querypacket (Fig. 3) and broadcasts it to check if there are any vehiclespresent behind it within the threshold. The broadcast packet,along with the HeaderID and the ClusterID, contains the position

and the direction information of the Header. This along with theTHDISTANCE allows the eligible vehicles to reply back to the Header.

In MC-DRIVE the threshold value THDISTANCE yields the optimallength of the cluster and is dependent on the speed and the radiorange of the vehicle approaching the intersection. This valueprovides the means for an effective cluster formation and cluster-head election and is computed as follows:

THDISTANCE ¼ ðLengthMaxþLengthMinÞ=2 ð1Þ

where,

LengthMAX ¼ RadioRange�9 MaxVelocity*TMINð Þ� MinVelocity*TMINð Þ9

LengthMIN ¼ RadioRange�9 MaxVelocity*TMAXð Þ� MinVelocity*TMAXð Þ9

TMIN ¼Distancef romStartPointtoEndPoint=MaxVelocity

TMAX ¼Distancef romStartPointtoEndPoint=MinVelocity

In the above equation, the factors TMIN and TMAX define theminimum and maximum time required by the vehicle to coverthe distance between Start Point and End Point. The End Point is areference point which signifies the termination of the clusterformation process when any vehicle of the cluster reaches it. TMIN

and TMAX along with the minimum and maximum velocities areused to compute LengthMAX and LengthMIN. These lengths repre-sent the distance between the elected clusterhead and thefarthest cluster member. These values give the maximum andminimum length of the cluster for the varying velocities andTHDISTANCE is a mean of these two values. Taking the mean valueas the threshold allows vehicles, irrespective of the speed, to be apart of a cluster until they reach the intersection.

Continuing with the process of clustering, when the Headerreceives replies from all the eligible vehicles within a stipulatedtime, it elects the farthest vehicle from itself which is belowTHDISTANCE as the clusterhead. However, if it does not receive anyreply, it checks for the existence of any vehicles within THDISTANCE

at every reference point marked along the road. In the process, ifthe Header reaches the Threshold Point and no clusterhead hasbeen elected then it elects itself as clusterhead. The Threshold

Algorithm:

If Start Point == TRUE && Cluster ID = =NULL then

{

Assign Vehicle = Header

Gen: Generate and send Query Packet and wait until timer expires

If reply received && Distance < THDISTANCE then

{

Elect clusterhead: Vehicle farthest from header and within THDISTANCE

}

Else

{

If Header is at Threshold Point

Elect self as clusterhead

Else

Goto Gen

}

}

Fig. 4. Algorithm for clusterhead election and cluster formation.

N. Maslekar et al. / Journal of Network and Computer Applications 36 (2013) 1308–1315 1311

Point is computed using THDISTANCE and is given by:

ThresholdPoint¼ StartPoint�THDISTANCE ð2Þ

The Threshold Point signifies the termination of the clusterheadelection procedure. Beyond this point electing the clusterheadwill not render a stable cluster. This ensures that irrespective ofthe speed, clusterhead will be a part of the cluster throughout thecluster lifetime. However, this is possible with an assumption thatthe start point is below the radio-range of the communicationvehicles.

The algorithm for the election of the clusterhead in MC-DRIVEis presented in Fig. 4.

In order to add members to the cluster, the clusterheadgenerates a query packet when it arrives at every reference point.This enables any new vehicles within THDISTANCE to join thecluster and also helps to maintain the existing cluster. This querypacket is similar to the query packet generated by the Headerwith an additional field which contains the list of vehicles presentin the cluster.

On receiving the query packet each vehicle checks if it istravelling in the same direction as the clusterhead and verifies ifits ID is present or not. If a vehicle detects that its ID does notexist, it checks if it is within THDISTANCE. If this condition isvalidated, it joins the cluster otherwise it ignores the message.

In both approaches, C-DRIVE and MC-DRIVE, the cluster isdismissed as soon as it crosses the intersection. As mentionedabove the clusterhead is responsible for computing and transmit-ting the information to the traffic signal control. The estimateddensity information within the cluster forms the platform toimplement the adaptive traffic controls named car–car commu-nication based adaptive traffic signal systems (CATS). The follow-ing section describes the computations and the implementationof this system.

4. CATS: Car–car communication based adaptive traffic signal

Prior to computing the adaptive cycle time, we need todetermine some parameters which will render a safe system.These parameters and modifications are described below.

4.1. Traffic light signaling terms

The basic timing elements within each phase of traffic signalcontrol include the green interval, the red interval, and inter-green interval. Here, a phase is defined as a distinct time period ofthe green, red and inter-green sequences. The green interval is theperiod of the phase during which the green signal is illuminated.The red interval represents the time following the yellow intervalin which all of the intersection’s signals are red. Finally, inter-green interval is the time between the end of green for one phaseand the beginning of green for another phase. It is the sum of theyellow and all-red intervals.

4.2. Inter-green interval

The inter-green interval consists of the yellow interval of onephase and red interval of the other phase. This time allows thevehicles that are already beyond the point-of-no-return to con-tinue through the intersection safely. To be precise, if the inter-green time is too short, only those vehicles that are close to theintersection will be able to continue through the intersectionsafely. In addition, only vehicles that are reasonably distant willhave adequate time to react to the signal and stop. The vehiclesthat are in between will be caught in the dilemma zone and won’thave enough time to stop or safely cross the intersection. Thus tohave a safe and efficient cycle time it is important to take intoaccount the design of the inter-green time. The modeling of this

Fig. 6. Packet transmitted to the traffic signal control.

N. Maslekar et al. / Journal of Network and Computer Applications 36 (2013) 1308–13151312

interval is specifically influenced by the safe stopping distance(SSD). The SSD is the minimum distance away from the intersec-tion which will enable the driver to safely stop a vehicle travelingat a specific speed without causing collisions at the intersection.

More specifically, the SSD is the distance during which thevehicle travels from the point at which the phase change isperceived to the time when the deceleration is complete. Thisdistance is the sum of lag distance and the braking distance. Thelag distance is the distance traveled by the vehicle during a periodcalled the reaction time which represents the time when thephase change is detected by the driver and the beginning of thedeceleration. Accordingly, the lag distance is computed by:

Lag Distance¼ 1:47vt ð3Þ

where, t is the reaction time, v is the velocity in m/s and 1.47 isthe constant to account for the driver’s eye height and the heightof the traffic signal (AASHTO, 2001).

The braking distance is the distance traveled by vehicle fromthe point when brakes are applied and to the point when thevehicle comes to a halt. This distance is computed as follows:

Braking Distance¼v2

30ðf þgÞð4Þ

where, v is the velocity of the vehicle, f the coefficient of frictionand g is the grade which provides information about the gradientof the road. Finally, the constant 30 is to account for the gravityfactor (AASHTO, 2001).

Hence, the safe stopping distance is determined by:

SSD¼ 1:47vtþv2

30ðf þgÞð5Þ

Based on the computation of SSD, the inter-green intervalwhich will allow the vehicles to continue through the intersectionsafely is given by:

T ¼SSDþLþW

1:47vð6Þ

where, L is the length of the vehicle and W represents the width ofthe intersection.

These mathematical models allow the computation and theimplementation of the cycle time for the proposed adaptivesystem. Cycle time and safety parameter computation togetherconstitutes CATS which obtain the density information fromvehicular networks. The architecture of CATS is shown in Fig. 5

4.3. Operations for adaptive traffic lights

As explained above, based on inter-vehicular communication,the density information for the vehicles approaching an

Fig. 5. Architecture of CATS.

intersection is computed at the clusterhead. Once the clusterheadis within the radio range (RR) of the traffic signal control ittransmits the density information to it.

The packet transmitted to the traffic signal control for thedensity information is shown in Fig. 6. Along with the directioninformation, this packet contains details of lane, arrival time andcluster length. The direction field, which is obtained through theclusters, gives information about the direction taken by a vehicleafter crossing the intersection and the lane field specifies the lanein which the vehicle is travelling. The other two fields, ClusterLength and the Arrival Time aid in formulating and implementingthe cycle time effectively.

In CATS, an optimum cycle length is defined based on the sumof lost time and the ratio of density to cluster length. Lost time isdefined as the sum of the inter-green time for all the phases oftraffic signal control. It is termed as lost time because no vehicleswill cross the intersection during this interval. The other twoparameters, namely the density and the cluster length, areobtained from the packets received from the clusters travellingin different directions. Based on these parameters, the cycle timeis formulated by modifying Webster’s equation (Webster, 1958).Webster’s model cannot be directly adopted for this adaptivesystem because the cycle time computation is based on designflow rate and saturation flow rate. These two parameters arecomputed based on historic data and hence it cannot be adaptedto varying demand. Thus the adaptive cycle length is given by:

Cycle Time¼1:5ltþ5

1�PðD=LnÞ

ð7Þ

where, lt is the sum of lost time usually taken as the sum of allinter-green periods (s), D/Ln is the ratio of density D in the clusterto the length Ln of the cluster.

The constants 1.5 and 5 in the formula are adopted from theanalytical model of Webster’s equation. To incorporate the degreeof error in the density, the cycle time is approximated to thenearest integer divisible by 5. This approximated value isobtained by experimental analysis. During this analysis, themaximum number of vehicles that can cross an intersectionwithin one cycle is verified. Once the cycle time is computed,the green splits for each lane approaching the intersection areallotted based on the D/Ln ratio. The smaller the value, higher isthe priority to allocate the green time. This is because a smallvalue indicates the requirement of more time for the cluster toclear the intersection. This strategy allows clearing most vehiclesat the intersection within the set cycle time and helps to reducethe average waiting time for vehicles. However, to make themodel more robust, the arrival time of the cluster at the inter-section is also taken into consideration. The inclusion of arrivaltime enables better decisions on when and which phase should beturned green so that the average waiting time is reducedconsiderably.

The simulation scenario and accuracy in density estimation forthe clustering algorithms, along with the gain of the proposedmodel in terms of a reduction in waiting time are discussed in thefollowing section.

5. Simulations and results

Through adaptive traffic lights, the objective is to obtain acollision-free system which reduces the average waiting time forvehicles at intersections. In the proposed model, a clustering

Fig. 8. Accuracy in density estimation.

N. Maslekar et al. / Journal of Network and Computer Applications 36 (2013) 1308–1315 1313

algorithm for VANETs is utilized to gather the density informa-tion. Hence along with the reduction in waiting times it becomesimperative to ascertain whether the gathered information isaccurate and timely with acceptable processing delays. To eval-uate and verify these factors through simulations, the NCTUnssimulation tool (Wang and Lin, 2008) is used.

For the purpose of analysis a topology of 3000 m�3000 mwith 7 intersections (Fig. 7) and a bi-directional road wereassumed. The simulator defines the mobility for vehicles whichfollow the designed road. For the mobility of each vehicle, arandom uniform distribution of speed was specified amongst thevehicles within a range from 8.5 m/s to 15 m/s.

To estimate the density we concentrate our analysis on thearea of twice radio range approaching the intersection. This valueis derived from an experimental analysis. It allows for retransmis-sion when the traffic signal control fails to receive the densityinformation. Table 1 summarizes the key parameters used for thesimulation.

To test the accuracy of the system, the simulations arerepeated for 15 times for each scenario and the average value ofthe obtained results is calculated. All the results depict theaverage number of vehicles that will be present in any givenlane. The presented results are the normalized values for the testswithin each case with a confidence interval of 95%.

Fig. 7. Simulation area.

Table 1Simulation scenario.

Quantity Value

Simulation time 100–1000 s

Vehicle speed 8–15 m/s

Transmission range 350 m

Number of intersections 7

Number of vehicles 5–40 (on each lane)

Mobility model Car following with overtaking

Mac Protocol 802.11p

Length of car 5 m

Distance between Reference Points 50 m

Start Point 300 m from the intersection

End Point 50 m from the intersction

THDISTANCE 180 m

Threshold Point 120 m from the intersection

Driver reaction time 2.5 s

Coefficient of friction 0.35 and 0.5

Grade 0.2

Width of intersection 20 m

Cycle time for pre-timed system 30 s

Number of simulations for each case 15

Since the working of CATS is based on density estimated by aclustering algorithm, it is necessary to compute the inaccuracythat occurs in the estimation of vehicle density. In this context,the term accuracy is defined as the ratio of the actual number ofvehicles present within the cluster to the number of predictedvehicles in the simulation scenario.

From Fig. 8, it is observed that in C-DRIVE and MC-DRIVE theaccuracy ranges from 89% to 99%. The inaccuracy in C-DRIVEresults from the clusterhead changes that are prevalent due tovehicles overtaking each other. Such effects are eliminated in MC-DRIVE because the cluster formation is governed by THDISTANCE.This allows for vehicles to be a part of a cluster until they reachthe intersection. A slight inaccuracy observed in MC-DRIVE is dueto the packet loss that occurs during the formation of clusters.This loss is the result of delivery failure in the join request at theclusterhead.

The achieved accuracy in density estimation yields an optimalcycle time for CATS and the system is compared with an adaptivefuzzy logic system (Kulkarni and Waingankar, 2007) and a pre-timed signal system. The cycle time for the pre-timed system isset by the simulator to 30 s for each phase. Under similarsimulation conditions it is found that CATS considerably reducesthe average waiting time experienced by vehicles at intersections.This is depicted in Fig. 9.

From Fig. 9 it is evident that as the density of vehiclesincreases, the average waiting time is increased. However, inthe proposed model, the increase is linear in behaviour ascompared to the exponential behaviour of the pre-timed systems.This is because during each cycle most of the vehicles are clearedthrough the intersection. Whereas in the case of a pre-timedsystem vehicles wait at the intersection for more than one cycle,hence gives exponential waiting time behaviour. In comparisonwith the adaptive fuzzy logic system, both C-DRIVE and MC-DRIVE render better performance as well. This is because thefuzzy logic system is dependent on a pre-defined rule. Therefore itcannot adapt the cycle timings to unexpected changes in thevehicular flow. In contrast, the proposed system uses cycletimings that are impromptu hence it can handle such situations.In the presented approach, the cycle time is set for vehicles whichare at a maximum distance of 300 m from the intersection. So, themaximum number of vehicles on a lane shows a saturationscenario with inter-vehicular distance of less than 4 m. Furtherincreasing the number of vehicles will not be practical, becauseit will lead to vehicular collisions. Thus, the obtained linearbehaviour is prevalent in traffic jam situations as well.

Fig. 9. Average waiting time for cars at the intersection.

Fig. 10. Average waiting time with and without arrival time consideration.

Fig. 11. Number of cars stopping at the intersection.

Fig. 12. Data convergence time.

N. Maslekar et al. / Journal of Network and Computer Applications 36 (2013) 1308–13151314

As discussed above, in the presented model the green timeswere allotted considering the arrival time of the cluster at theintersection. To verify the effect, of arrival time on the system, themodel was tested without considering the arrival time (Fig. 10).This arrival time is obtained from the packet transmitted byclusterhead (Fig. 6).

It is evident from the results that the arrival time providesbetter decisions on when a phase should be turned green so thatthe average waiting time is reduced considerably.

In terms of efficiency, as shown in Fig. 11, it is found that CATSreduces the number of vehicles that stop at the intersection byalmost 20% for C-DRIVE and 12% for MC-DRIVE. This is becausewithin the computed cycle the optimal green splits to differentphases are allotted based on the arrival time of the cluster. Thisreduction showed a decreased queue length at the intersection.

In the simulation results, it is observed that the performance ofMC-DRIVE is slightly inferior for CATS especially for the numberof vehicles that stop at the intersection. This is because inC-DRIVE the clusterhead is at the front of the cluster whereas inMC-DRIVE it is at the center of the cluster. Thus when theinformation is transmitted to the traffic signal control, thedistance to be travelled by the first vehicle in the cluster isgreater in C-DRIVE than in MC-DRIVE. For example, in the worstcase scenario in MC-DRIVE, the distance between the clusterheadand the first vehicle might be equal to the radio range. In such acase, when the density information is transmitted to the trafficsignal control the first vehicle in the cluster will be at theintersection and will be held up in the previous cycle. Thus,based on the phase timing of the previous cycle the first vehiclemay either cross or wait at the intersection. Also in the case ofretransmission of density information, some nodes belonging to

the cluster may reach the intersection and may get held up inprevious cycle timings thereby increasing their waiting timeat the intersection and resulting in a degraded performance forMC-DRIVE in CATS.

From the results, it is apparent that CATS shows an improve-ment in terms of average waiting time and the percentage ofvehicles stopping at intersections. However these results are notadequate to validate the model. To ascertain the advantage of theproposed adaptive system it is necessary to analyze the proces-sing time to convert the received density into efficient cycle timeand green splits. This is particularly important because it involvescommunication between vehicles.

Through simulation and analysis it is observed that the timerequired to process the data is less than 200 ms for C-DRIVE andaround 250 ms for MC-DRIVE (Fig. 12).

The convergence time in MC-DRIVE is greater because thetraffic signal control has to validate the arrival time of the cluster.This is necessary because, as mentioned above, some nodesbelonging to the cluster may reach the intersection and may getheld up in previous cycle timings. Hence the new cycle computa-tion should reduce the arrival time accordingly. In C-DRIVEbecause of the simple clusterhead election policy, these retrans-missions do not affect the convergence time.

However, in both approaches, the obtained time is quite smallwhen compared to the time taken for the entire cluster to reachthe intersection. Hence the obtained value validates the protocolin terms of data convergence time.

Since VANETs are used to gather density information, it is alsonecessary to consider the latency of the communication betweenvehicles and the infrastructure. In CATS this delay is around210 ms which is well within the limits. Finally, one importantaspect is the fact that no vehicular collisions occurred at the

N. Maslekar et al. / Journal of Network and Computer Applications 36 (2013) 1308–1315 1315

intersection. During the simulations it is observed that for 60simulations representing the various cases and 7 intersections thesystem was collision free. This shows that the proposed system isa safe model. This is because the cycle time calculation includesthe inter-green time and the safe stopping distance.

6. Conclusion and future work

In this paper an adaptive traffic signal system based on car-to-car communication was presented. Two approaches were adoptedand compared for density estimation at intersections. The simula-tion results demonstrate that with the presented model a betterlevel of service (LOS) (Idaho et al., 2003) in terms of the averagewaiting time at the intersections is achieved. The LOS forsignalized intersections is defined in terms of average stoppeddelay per vehicle. In the literature five LOS categories A to E arementioned, with A being the best and E the worst. The averagedelay for category A is less than 5 s and for E it is more than 40 s.With reference to these categories of LOS, the proposed adaptivesystem CATS fits between categories A and B. Also the proposedadaptive system reduces considerably the length of a waitingqueue. Thus, with fewer vehicles waiting, this system improvesthe effective travelling time. From the point of view of computa-tion time, the proposed model is efficient enough to compute aneffective cycle time within a given time frame. It was alsoobserved that CATS performed better with a simpler approachof C-DRIVE compared to MC-DRIVE. Hence it will be appropriateto conclude that adaptive traffic lights based on car to carcommunication has a good potential for improving the trafficconditions in urban areas.

As a part of further development, we would like to develop theprototype of the presented solution and test its performance in apractical scenario. Also, the existing two clustering algorithms areprimarily based on the direction information which is obtainedvia GPS and digital maps. With digital maps and GPS, the directiondetection becomes relatively easy to manage. However, to makethe algorithms robust to situations when these inputs are notavailable it is necessary to investigate and design a relativedriving direction detection algorithm.

Acknowledgement

The authors would like to thank the irseem-Esigelec forfinancially supporting this research under the Geocolis project.

References

AASHTO. (2001). Standard specification for structural supports for highway sign,

luminaries and traffic signals. 4th ed. American Association of State Highwayand Transportation Officials.

Amin, M, & Jalili, S (2006). Urban signal control using intelligent agents. 7thinternational FLINS conference on applied artificial intelligence, (pp. 811–816).Genova, Italy.

Dresner, K, & Stone, P (2004). A reservation-based intersection control mechanism.3rd international conference on autonomous agents and multiagent systems,

(pp. 530–537). New York, NY, USA.Eichler, S (2007). Performance evaluation of 802.11p WAVE Communication. 1st

IEEE international symposium on wireless vehicular communication, (pp.2199–2203). Baltimore.

Gradinescu, V, Gorgorin, C, Diaconescu, R, Cristea, V, & Iftode, L (2007). Adaptivetraffic ligths using car-to-car communication. 65th IEEE vechicular technologyconference, (pp. 21–25). Dublin, Ireland.

Gregoire, P, Desjardins, C, Laumonier, J, & Chaibdraa, B (2007). Urban traffic controlbased on learning agents. 10th international IEEE conference on intelligent

transportation systems, (pp. 916–921). Seattle, Washington, USA.Huang, Q, & Miller, R (2004). Reliable wireless traffic signal protocols for Smart

intersections. San Antonio, Texas: 14th annual meeting of ITS America.Idaho, U, Oregon State, U, & Portland State, U (2003). Retrieved 2010, from

Transportation Engineering Online Lab Manual: /http://www.webs1.uidaho.edu/niatt_labmanual/S.

Kulkarni, G, & Waingankar, P (2007). Fuzzy logic based traffic light controller. 2ndinternational conference on industrial and information systems, (pp. 107–110).Penadeniya.

Kutil, M, Hazalek, Z, & Cervin, A (2006). Balancing the waiting times In a SimpleTraffic Intersection Model. 11th IFAC symposium on control in transportation

systems. The Netherlands.Lammer S, & Helbing D. Self-control of traffic lights and vehicle flow in urban road

networks. International Journal of Statistical Mechanics: Theory and Experiment2008:1–36.

Liu Z. A survey of intelligent methods in urban traffic signal control. InternationalJournal of Computer Science and Network Security 2007;7:105–12.

Maslekar, N, Boussedjra, M, Labiod, H, & Mouzna, J (2009). Direction based

clustering algorithm for data disseminaiton in VANET. 1st IEEE vehicularnetworking conference, (pp. 1–6). Japan.

Maslekar, N, Boussedjra, M, Labiod, H, & Mouzna, J (2011). C-DRIVE: clusteringbased on direction in vehicular environment. 4th IFIP international conference

on new technologies, mobility and security, (pp. 1–5). Paris.Mizuno, K, Fukui, Y, & Nishihara, S (2008). Urban traffic signal control based

on constratin satisfaction. 41st international conference on system sciences,(pp. 1–10). Hawaii.

Wang, S, & Lin, C (2008). NCTUns 5.0: a network simulator for IEEE 802.11p and

1609 wireless vehicular network researches. 68th IEEE vehicular technologyconference, (pp. 1–2). Calgary, Alberta.

Webster F. Traffic signal settings. Road research laboratory technical report.London: HMSO; 1958.

Wiering M, Veenen J, Vreeken J, & Koopman A. Intelligent traffic lightcontrol.Universiteit Utrecht, Dept. of Information and Computing Sciences. Utrecht:Universiteit Utrecht; 2004.