A Decentralized Network with Fast and Lightweight Autonomous ChannelSelection in Vehicle Platoons for Collision Avoidance

Ankur Sarker, Chenxi Qiu, and Haiying ShenDepartment of Electrical and Computer Engineering

Clemson University, Clemson, USA{asarker, chenxiq, shenh}@clemson.edu

Abstract—Always keeping a certain distance between ve-hicles in a platoon is important for collision avoidance.Centralized platoon systems let the leader vehicle deter-mine and notify the velocities of all the vehicles in theplatoon. Unfortunately, such a centralized method gener-ates high packet drop rate and communication delay dueto the leader vehicle’s limited communication capability.Therefore, we propose a decentralized platoon network,in which each vehicle determines its own velocity byonly communicating with the vehicles in a short range.However, the multiple simultaneous transmissions betweendifferent pairs of vehicles may interfere with each other.Directly applying current channel allocation methods forinterference avoidance leads to high communication costand delay in vehicle joins and departures (i.e., vehicledynamics). As a result, a challenge is how to reduce thecommunication delay and cost for channel allocation in de-centralized platoon networks? To handle this challenge, byleveraging a typical feature of a platoon, we devise a chan-nel allocation algorithm, called the Fast and LightweightAutonomous channel selection algorithm (FLA), in whicheach vehicle determines its own channel simply based onits distance to the leader vehicle. We conduct experimentson NS-3 and Matlab to evaluate the performance of ourproposed methods. The experimental results demonstratethe superior performance of our decentralized platoonnetwork over the previous centralized platoon networksand of FLA over previous channel allocation methods inplatoons.

1. Introduction

Vehicle platoon systems, as a type of next-generation of land transportation systems, have receivedmuch attention during recent few years. In a platoon,one leader vehicle and several follower vehicles drivein a single lane, where each vehicle maintains a dis-tance from its preceding vehicle. Since the platoonsystem allows for a shorter distance between vehicles,it provides higher traffic throughput and better trafficflow control [1], [2]. In addition, it helps to reduceenergy consumption by avoiding unnecessary changesof acceleration. However, the shorter distance betweenvehicles leaves less time for each vehicle to react when

(a) Centralized platoon network

(b) Decentralized platoon network

Figure 1. Centralized vs. Decentralized platoon networks.

its preceding vehicle decreases speed, which may causecollisions and impair vehicle safety. Therefore, it is crit-ical to build a well-connected communication networkfor a platoon so that vehicles can quickly adjust theirvelocities through fast communication.

Several centralized platoon systems [3]–[6] havebeen proposed, where the leader vehicle is exclusivelyresponsible for maintaining the entire platoon. Thatis, the leader vehicle determines the velocities andthe trajectories for all follower vehicles and notifiesthem by direct communication. Also, most of thesesystems use static formation and do not allow vehiclejoins or departures during run time. Fig. 1(a) showsa centralized platoon system, where the leader vehi-cle sends the information to three follower vehiclesdirectly. It means that the leader vehicle plays a criticalrole and it must be capable of communicating withall vehicles. Usually, the leader vehicle is equippedwith a special communication device, e.g., the DSRC(Dedicated Short Range Communication) based wire-less device, to communicate with other vehicles. Thelimited communication range of such devices (i.e., 300-500 meters [7]) constrains the active platoon length (i.e.,the number of vehicles in the platoon). For example, ina centralized platoon system, when the speed limit is 20meters/second and the safety distance between vehiclesis 35 meters according to the traffic policy [8], then theplatoon can only support b300/35c−b500/35c = 8−14vehicles. Moreover, during the running time, a highervelocity increase requires longer inter-vehicle distancefor collision avoidance but leads to fewer vehicles thatcan be connected with the leader vehicle. In addition,the long distance between the leader vehicle and oth-

er follower vehicles significantly increases the packetdrop rate in communication due to signal’s propagationlosses [9] and multipath fading effect [10], especiallyin the urban transportation environment. Furthermore,to avoid contentions between follower vehicles, thecommunication procedures between the leader and fol-lower vehicles are scheduled separately, which leadto high packet delay in the network. Hence, due tohigh packet drop rate and delay in communicationnetwork, the vehicles’ safety cannot be guaranteed inthe centralized platoon networks. In current centralizedplatoon systems, IEEE Wireless Access in VehicularEnvironment (WAVE) is based on DSRC technologyand it defines the architecture and services for multi-channel DSRC/WAVE devices. WAVE combines IEEE802.11p and IEEE 1609 protocol suite, covering fromphysical layer to application layer. On the physicallayer, DSRC is operated at 5.9GHz band and there areseven channels for control and service operations. Onthe MAC layer, DSRC extends the basic service setwith the Enhanced Distributed Channel Access (ED-CA) mechanism for classifying different data flow intodifferent access categories. To deal with the hidden ter-minal problem, IEEE 802.11p utilizes the carrier sensemultiple access with collision avoidance (CSMA/CA)mechanism, the signal overhead on the control channeldue to the handshaking procedures could penalize thesafety-critical messages, especially in platoon systemsand which may show poor performance with heavypacket loss and average delay in coarse traffic scenario[11]. Besides, CSMA/CA may exhibit exposed terminalproblem for vehicles inside platoon. Moreover, DSRCis pushed to the limit if different application scenar-ios (e.g., information and warning function, longitudecontrol, safety, cooperative assistance, etc) which createcontradictory constraints under heavy traffic condition.To overcome such drawbacks of the centralized platoonnetworks, we propose a decentralized platoon networkwith the objective to guarantee individual vehicle safetyand increase the number of vehicles in a platoon. Weconsider that each vehicle is equipped with a mobilecommunication device capable of communicating with-in a short distance (e.g., IEEE 802.11a/b enabled mobiledevice which covers 70m-80m [12]). In our proposeddecentralized platoon network, each vehicle is onlyrequired to communicate with its neighbor vehicles andthere is no explicit centralized control. Also, all vehiclesare independent and can join in or leave from theplatoon any time. As shown in Fig. 1(b), each vehicleperiodically transmits a message composed of its ownvelocity and location to the next vehicle in the platoon,and each vehicle calculates its own velocity based on itscurrent velocity and received information from its pre-ceding vehicle. As the leader vehicle does not need tocommunicate with all the vehicles, the platoon length isno longer limited by the leader vehicle’s communicationcapability. Besides, the decentralized network has muchlower packet drop rate since each vehicle only needs tocommunicate with the vehicles in a short range. Further-

g segments (group)

... ...

segment

g segments (group)

Figure 2. Partition of a platoon.

more, different from the centralized platoon networks,the decentralized network allows multiple transmissionsto be active simultaneously, which reduces the packetdelay. Therefore, the decentralized network is moreeffective for collision avoidance.

However, the decentralized platoon network bringsabout the transmission interference problem. Differentfrom the centralized systems, where only one transmis-sion (between the leader vehicle and any other followervehicle) is allowed in a single time slot, the decen-tralized platoon network allows multiple transmissionsto be active simultaneously. The multiple simultaneoustransmissions between different vehicle pairs may in-terfere with each other. To avoid interference, we needchannel allocation to schedule different channels to ve-hicles. Using the previous channel allocation algorithms[13]–[18], to select a channel to use, a vehicle mustestimate the sum interference from all other vehicles,which requires the knowledge of their locations. Then,if some vehicles’ locations in the platoon are changedwhen a vehicle enters or leaves the platoon [19], theymust send messages to all other vehicles to update theirlocations, which generates high communication cost.Since vehicles in platoon may change their locations[19], directly employing the previous channel allocationalgorithm to the platoon network would lead to muchhigher communication cost and longer transmission de-lay. Considering the poor channel capacity for the vehi-cle to vehicle (V2V) communication [14], a challengeis how to conduct channel allocation with low delayand low communication cost in a decentralized platoonnetwork?

In this paper, we aim to resolve the channel in-terference problem arisen in the decentralized platoonnetwork. To handle this problem, we propose a Fastand Lightweight Autonomous channel allocation algo-rithm (FLA) that takes advantage of a typical featureof the platoon. Different from general wireless net-works, where the nodes are arbitrarily distributed, inthe platoon, vehicles drive in a single lane and thedistance between neighboring vehicles is equal to thesafety distance [8]. Based on this feature, to avoidinterference, we let vehicles use the same channel onlywhen their distance is beyond the interference range andlet vehicles within the interference range use differentchannels. Interference range is the distance range thatmakes the interference upper bounded by an acceptablevalue for packet decoding. Specifically, as shown in Fig.2, we geometrically partition the platoon into segmentsof the same length δ such that each segment contains at

most one vehicle and the length of a segment is lowerbounded by the interference range. Then, we considerevery g consecutive segments as a group of vehiclesand allocate g different channels to the segments ineach group of vehicles. Here, g is the current minimumnumber of channels needed to avoid the interference.The aforementioned platoon feature also enables a ve-hicle to locate its segment position in a group andthen autonomously decide its channel accordingly. Asa result, the vehicles using the same channel have adistance equals to the interference range in between.Finally, we evaluate our decentralized platoon networkthrough simulation in both Matlab [20] and NetworkSimulator 3 (NS-3) [21]. The simulation results demon-strate the superior performance of our decentralizedplatoon network over the traditional centralized platoonnetwork in both aspects of packet drop rate and platoonsize. The simulation results also show the efficiency ofour channel allocation method compared to the previousmethods [16], [22]. The rest of this paper is organized asfollows. Section 2 presents the aforementioned channelinterference problem in detail. Section 3 presents ourmethods to solve the previously mentioned problemand Section 4 evaluates the performance our methods.Section 5 presents the related work. Section 6 concludesthis paper with our remarks on the future work.

2. Problem Statement

In this section, we first discuss the channel al-location problem in wireless communication network.Then, we introduce the channel allocation problem invehicular platoon network.

2.1. Background

We consider a platoon with n communication linksamong vehicles: (s1, r1), ..., (sn, rn), where (si, ri) rep-resents a transmission link from vehicle sender si tovehicle receiver ri. Let S = {s1, ..., sn} and R ={r1, ..., rn} represent the set of vehicle senders andvehicle receivers. Let xsi and xridenote the location ofvehicle sender si and vehicle receiver ri, respectively.We calculate the Euclidean distance between si and rjby dsi,rj = |xsi − xrj |, ∀si ∈ S and ∀rj ∈ R. Weassume that the mobile devices in vehicles have thesame communication range R and any sender, say si,can communicate with its receiver ri iff dsi,ri ≤ R.We use δ to denote the sum of the safety distance andvehicle’s general length, called vehicle distance. Here,we assume all vehicles have the same length. That is, ifthere are m vehicles in the platoon, the platoon length isapproximately δm. Accordingly, we require dsi,rj ≥ δ,∀si ∈ S and ∀rj ∈ R for collision avoidance. Table 1presents the main notations used in this paper. As weindicate in Section 1, the decentralized platoon networkcan overcome the drawbacks of the centralized platoonnetwork by increasing the platoon length and vehicle

s1

useful signal useful signal

interferencer1 s2 r2

Figure 3. Vehicle signal interference.

safety, but it also brings the interference problem. In thefollowing, we will explain the problem in more detail.

In the centralized platoon network, only one trans-mission is allowed in a single time slot between theleader vehicle and a follower vehicle. The decentralizedplatoon network allows multiple transmissions betweeneach pair of consecutive vehicles to be active at thesame time, which reduces packet drop rate caused bylong communication distance and packet transmissiondelay. However, when multiple vehicles transmit theirpackets at the same time with the same channel, theirsignals interfere with each other. As Fig. 3 shows,vehicle s1 sends a packet to vehicle r1, and vehicles2 sends a packet to vehicle r2. Because vehicle r1 iswithin the communication range of vehicle s2, vehicler1 will be interfered by vehicle s2, which may corruptthe packet received from vehicle s1. Therefore, weneed to allocate different channels to vehicles withinthe interference range to overcome interference amongmultiple transmissions.

TABLE 1. NOTATIONS AND DEFINITIONS.

Symbol Descriptionsi The vehicle sender iri The vehicle receiver iSINRsi,ri SINR received by ri from vehicle siN0 Noise power received by each vehicle(si, ri) The link between si and riP The transmission powerR The transmission rangen The number of linksδ The sum of the safety distance and vehicle’s

general lengthS The set of sendersR The set of receiversU(i) The set of vehicle senders with the same

channel as vehicle sender siα Path loss exponentγth The decoding thresholddsi,rj The Euclidean distance between vehicles si

and rjxsi (xri ) The location of si (ri)

Since the interference among transmissions is themain factor to be considered for channel allocation,the selection of the interference model is important.In this paper, we select the Signal-to-Interference-plus-Noise Ratio (SINR) model [13], [23] to describe theinterference among transmissions, which is more real-

istic than the traditional graph-based interference model[16], [24]–[26]. The SINR received by receiver ri fromnode si is defined as:

SINRsi,ri =Pd−αsi,ri

N0 +∑∞

sj∈U(i)\si Pd−αsj ,ri

, (1)

where U(i) represents the set of vehicle senders thatuse the same channel as vehicle sender si, P representsthe transmission power, N0 represents the whole noisepower, and α represents the path loss exponent, whichreflects the reduction of signal power as the signalpropagates through space. The SINR model considersthe interference received by each vehicle receiver, sayri, as the sum of all of its received noise (N0) andundesired transmissions’ signals (

∑∞sj∈U(i)\si Pd

−αsj ,ri).

ri can successfully receive the packet from vehicle si iffits received SINR is higher than the decoding thresholdγth.

Equ. (1) shows that the SINR received by eachreceiver ri is determined not only by the distancebetween ri and its sender vehicle si (i.e., dsi,ri), butalso by the distance between vehicle ri and all othervehicles that share the same channel with vehicle si(i.e., dsj ,ri , sj ∈ U(i)\si). Therefore, to determinewhether a channel can be allocated to si, a straight-forward channel allocation algorithm [13] is to collectthe location information of all other vehicles, estimatetheir interference to vehicle ri according to Equ. (1) andsee whether the SINR received by vehicle ri is higherthan the decoding threshold γth. If the SINR is lowerthan γth, the transmission will be failed and vehiclesi cannot use this channel. One the other hand, if theSINR is higher than γth, then the transmission will besuccessful and vehicle si can use this channel.

If we apply this approach to the platoon network,the leader node needs to function as the central nodeto calculate the channel allocation and notify all nodesabout their allocated channels. However, this approachis impractical to be applied in the platoon networkbecause vehicles in platoon may change their locationsin vehicle dynamics, leading to the changes of distancesbetween vehicles. The high computation time complex-ity of the recalculation of the channel allocation of allnodes leads to a long delay and the notification of allo-cated channels to all nodes leads to high communicationcost and also communication delay.

2.2. Vehicle Channel Allocation Problem

Below, we first formulate the Vehicle Channel Allo-cation (VCA) problem.

Formally, the VCA problem is defined as follows:Instance: A finite set of senders S and their respectivereceivers R in a geometric plane, decoding thresholdγth, and a constant Λ.Question: Using Λ channels, whether there exists aschedule (which allocates a channel to each vehiclesender), such that the SINR received by each vehiclereceiver is higher than γth?

3. System Design

In this section, we introduce how to solve the Vehi-cle Channel Allocation (VCA) problem using a heuristicmethod in the decentralized platoon network.

As mentioned previously, vehicles may change theirlocations in the platoon. Such changes also lead to in-terference to change under the given channel allocation,which requires us to keep re-scheduling channels amongvehicles whenever location change happens. PreviousSINR based channel allocation methods [13]–[15] re-quire each node to collect the location information ofall the vehicles, which leads to high transmission delay,especially when re-scheduling of channels frequentlyhappens. In addition, high computation time complexityof the recalculation of the channel allocation of all n-odes leads to high computation delay. To solve the prob-lem, we propose a Fast and Lightweight Autonomouschannel allocation algorithm (FLA), where each vehicleautonomously determines its channel for transmissionsolely based on its distance from the leader vehicle,without the need to collect the location information ofother vehicles. The idea of FLA is based on the platoonfeature that the distance between vehicles equals thesafety distance. Based on this feature, we can conductthe channel allocation to ensure that the distance be-tween vehicles using the same channel is lower boundedby the interference range. Also, each vehicle can au-tonomously decide which channel it should use solelybased on its distance from the leader vehicle, wherethe distance with the location of the leader vehicle isbroadcasted from the leader vehicle. Here, the leadervehicle needs to periodically broadcast its location tothe following vehicles. According to this informationand its own location, each following vehicle can deriveits distance from the leader vehicle. As Fig. 4 shows,we geometrically split the platoon to g segments withlength equal to δ so that each segment contains at mostone vehicle. Then, we consider every g consecutivesegments as a group. Next, we allocate g channels to gsegments in each group, and the vehicle in a segmentchooses the channel allocated to this segment. As aresult, at most one vehicle is contained in each segmentand the vehicles sharing the same channel must havedistance no less than the interference range, (i.e., kgδ,k = 1, 2, ...), which avoids the interference. In the fol-lowing, we will introduce how to determine g which isthe minimum number of channels to avoid interference(Section 3.1) and how each node determines its channelin FLA (Section 3.2) in detail.

3.1. The Minimum Number of Channels

Now, we need to determine g, which denotes theleast number of channels used to overcome interference.For any segment l, the distance between segment l andeach segment that has the same channel as segment l iskgδ (k = 1, 2, ...). If the distance between two segmentsis kgδ, then the safety distance between the vehicles in

... 1 ... 1

segment size = d

g different channels

g 2 g 2

the distance = gd

Figure 4. Channel allocation.

the two segments is kgδ − δ, which implies that theinterference generated from the vehicle in one segmentto the vehicle in the other segment is at most P (kgδ−δ)−α. Consequently, the sum interference received byeach vehicle is upper bounded by∞∑k=1

P (gkδ − δ)−α ≤ P

∞∑k=1

((g − 1)kδ)−α

≤ P (g − 1)−αδ−α∞∑k=1

k−α

= P (g − 1)−αδ−αζ(α) (2)

where ζ(α) =∑∞

k=1 k−α. By definition (Equ. (1)),

SINR is actually the quotient of the useful signal powerto the sum interference. Because the distance betweeneach pair of sender and receiver, say vehicles si andri, is upper bounded by the communication range R,the useful signal power received at vehicle ri (Pd−αsi,ri)is lower bounded by PR−α, i.e., Pd−αsi,ri ≥ PR−α.To guarantee that ri can successfully receive a packetfrom si, we need to ensure that SINRsi,ri ≥ γth. Thatis, we need to find a channel for si to upper boundthe sum interference from the senders with the samechannel with si by PR−α

γth. Then, according to Equ. (2),

we need to ensure

P (g − 1)−αδ−αζ(α) ≤ PR−α

γth(3)

from which we derive that

g ≥ d(Rαδ−αζ(α)γth

) 1α + 1e (4)

We hope we can use as fewer channels as possible, then:

g = d(Rαδ−αζ(α)γth

) 1α + 1e (5)

That is, g can be pre-defined based on the transmissionrange of vehicles (R), path loss exponent (α), decodingthreshold γth, and segment distance δ.Theorem 3.1. (Feasibility) By setting the number of

channels g by Equ. (5), the SINR received by eachreceivers is higher than the decoding threshold γth.

Proof: Without loss of generality, we examineany vehicle receiver ri of which (si, ri) ∈ Lk. Because(si, ri) ∈ Lk, 2k−1δ ≤ dsi,ri < 2kδ, the signal powerreceived at ri from its desired sender si is at least

Psi,ri ≥P

2αkδα. (6)

Now we consider the interference caused by the trans-mission from other requests. Suppose ri is located in

square Segkm, since links are scheduled concurrently ifftheir receivers reside in the segment with the same color,the interference can only be caused by the links whosereceivers are in Segkm±2q, where q ∈ N. We representthe set of links whose receivers are in the two segmentsby Qkq . For any link (sj , rj) ∈ Qkq , because the distancebetween ri and sj is at least (2q(gk − 1) − 2k)δ, theuseful signal on ri is at most

Psi,ri ≤P

(2q(gk − 1)− 2k)αδα. (7)

and ∑sj :(sj ,rj)∈Lk\(si,ri)

Psj ,ri =

∞∑q=1

∑j:(sj ,rj)∈Qkq

Psj ,ri

Then,

SINRsi,ri =Psi,ri∑

sj :(sj ,rj)∈Lk\(si,ri) Psj ,ri(8)

≥ (gk − 1)α

2αk+1ζ(α)≥ γth. (9)

which implies that ri can successfully receive the pack-et.

3.2. Autonomous Channel Determination

In one segment group, as shown in Figure 4, eachsegment has a segment ID ranging from 1 to g. Vehiclesin the segment with ID i choose to use channel i amongg channels. Below, we introduce our FLA algorithmin which each vehicle autonomously determines itssegment ID and then the channel to use.Definition 3.1. (Distance offset). The distance offset of

a follower vehicle receiver ri, denoted by ∆i, isdefined as the remainder of its distance from theleader vehicle (r1) divided by gδ:

∆i = dri,r1 mod gδ (10)

Property 3.1. Given the distance offset of a receiver ri,∆i, the segment ID of this vehicle is

⌈∆i

gδ

⌉.

TABLE 2. THE FLA TABLE.

∆i [0, δ) [δ, 2δ) ... [(k − 1)δ, kδ)Channel 1 2 ... g

According to Property 3.1, each vehicle’s distanceoffset determines its segment ID, and then determinesits channel. Hence, we can build a table (Table 2),namely the FLA table, which associates each distanceoffset with each channel in g channels. A vehiclereceives this table from its preceding vehicle after itjoins the platoon. This table is kept in each vehicle’sstorage. Since the partition is static over time, once thetable is built, each vehicle does not need to change theFLA table anymore. Using the FLA table, each vehicle

only needs to know its distance from the leader vehicleto determine its channel without the need to collectlocation information of other follower vehicles. To letall the follower vehicles know the leader vehicle’s loca-tion, the leader vehicle’s current location is periodicallypropagated to all the follower vehicles by piggybackingthe location information on the packet periodically sentfrom a preceding vehicle to its succeeding vehicle.According to the leader vehicle’s location, each fol-lower vehicle can calculate its distance from the leadervehicle. To implement FLA, each vehicle only needs tocalculate the distance offset based on its current distancefrom the leader vehicle by Equ. (10). Then, it checksthe FLA table by the calculated distance offset andfinds the corresponding channel. For example, supposethe safety distance is 30 meters, and the number ofchannels is g = 5. Vehicle i estimates that the distancebetween the leader vehicle and itself is 195 meters.Using Equ. (10), vehicle i’s distance offset equals 195mod (30 × 5) = 45 meters. Since 45 ∈ [30, 60), itchooses channel 2 based on the FLA table.

4. Performance Evaluation

In the following subsections, we evaluate the per-formance of proposed distributed platoon network withseveral state-of-the art methods. For the simulation s-tudy, we use both Network Simulator-3 (NS-3) [21] andMatLab [20]. In the case of channel allocation method,we implemented several methods in Matlab and in thecase of platoon networks, we use NS-3 for evaluatingthe performance of different methods.

For the simulation, we fixed one leader vehicle withother thirty follower vehicles where each vehicle wascapable of communicating with its neighbor vehicleswithin a communication range. We considered the com-municator device was IEEE 803.11b enabled and itsrange was about 80m. Each vehicle was capable tochange its velocity from 8m/s to 30m/s (17.89m/h--67.12m/h), independently. In addition, each vehiclemaintained safety distance (47.5m–80m) based on thesafety policy [8]. Thus, based on the safety distancesand communication range of each vehicle, each vehiclewas capable of communicating with almost two neigh-bor vehicles. Here, the communication range of eachvehicle is larger than the one times of safety distance butsmaller than the two times of safety distance. Besides,each vehicle changed its velocity at every 0.1 secondif it was necessary. In the following, we compare ourmethod with previous methods in two aspects explainedbelow.1) Decentralized platoon network. We chose an exist-ing centralized vehicle network, called DynB [6], forcomparison. In DynB, the leader vehicle continuouslysends beacon messages to other vehicles. The leadervehicle requires a special device (802.11p DSCR de-vice) to communicate with all other vehicles, and itstransmission range is about 300–500 meters. We usedNS-3 for this test.

2) Channel allocation. We chose two channel allocationmethods, which are based on graph interference model[16] and SINR model [14] for comparison. Both meth-ods require the location information from all nodes.Using the method for the platoon application, whenallocating a channel, the method in [14] iteratively picksa sender and removes all the senders that have SINRsmaller than the decoding threshold. This process isrepeated until each sender is either picked or removed.In [16], for each pair of vehicles, a central node buildsan edge between the two vehicles iff they are withineach other’s transmission range. Then, the central nodepartitions all the nodes into several subsets, where thevehicles in each subset are not connected by any edge.Finally, the central node allocates a channel to eachsubset. In this test, we used Matlab to evaluate differentmethods.

4.1. Decentralized platoon network

We measured the average packet drop rate, whichis defined as the ratio of the number of packets droppedto the number of packets sent to all the vehicles ineach second. Fig. 5(a) shows the average packet droprates of the DynB centralized platoon network andour decentralized platoon network when the numberof vehicles in the platoon is varied from 6 to 30. Asexpected, our method has much smaller packet droprate than that of DynB. When the number of vehiclesin Platoon is 30, the packet drop rate of DynB isapproximately 0.8, whereas our method’s packet droprate is less than 0.37. It is because that DynB requiresthe leader vehicle to communicate with all the followervehicles, in which the message signal will fade signifi-cantly over a long distance [13]. While in our method,each vehicle is only required to communicate with itspreceding vehicle. A high packet drop rate seriouslyaffects the vehicle safety. Therefore, our decentralizedplatoon network can support higher vehicle safety.

We also compared the number of vehicles that canbe supported by the DynB centralized platoon networkand our decentralized platoon network. Here, if a ve-hicle’s packet drop rate (i.e., the ratio of its droppedpackets in its sent packets in each second) is lowerthan 0.3, we consider that this vehicle can be support-ed by the platoon. Fig. 5(b) compares the number ofvehicles that the centralized and decentralized platoonnetworks can support when the vehicle velocity is var-ied from 8m/s to 30m/s. We find that as the vehiclevelocity increases, the number of vehicles decreasesin the centralized platoon network, while it maintainsthe same level in the decentralized platoon network. Itis because that in the centralized platoon network, thetransmission range of the leader vehicle is limited butthe inter-vehicle safety distance increases as the vehiclevelocity increases [8]. Different from the centralizednetworks, in our decentralized platoon network, eachvehicle only needs to send the packets to its succeedingvehicle. Hence, as long as the inter-vehicle distance

0

0.2

0.4

0.6

0.8

1

6 10 14 18 22 26 30Av

e.

pa

ck

et

dro

p

rate

# of vehicles

Decentralized platoonCentralized platoon

(a) Average packet drop

0

10

20

30

40

8 13 18 23 28

# o

f v

eh

icle

s

Velocity (m/s)

Decentralized platoonCentralized platoon

(b) The number of vehicles

1.5

1.9

2.3

2.7

8 12 16 20 24 28

Av

e.

pa

ck

et

de

lay

(s

)

# of vehicles

Decentralized platoonCentralized platoon

(c) Average packet delay

0255075

100125150175

8 13 18 23 28Tota

l #

of

safe

ty

vio

lati

on

s

# of vehicles

Decentralized platoonCentralized platoon

(d) Safety violation

Figure 5. Comparison of the centralized and the decentralized platoon networks.

is smaller than the communication range, the platoonis always connected no matter how velocity changes.Consequently, the number of vehicles that platoon cansupport is not affected by the velocity changes. Thisresult confirms that our decentralized platoon can sup-port more vehicles than the existing centralized platoonnetworks. Fig. 5(c) compares the average packet delayof the decentralized and centralized platoon networks.We find that the average packet delay of the centralizedplatoon network is higher than that of the decentralizedplatoon network. This is because higher packet droprate (as shown in Fig. 5(a)) generates more retransmis-sions, which leads to higher packet delay. Finally, wecompared the total number of safety violations duringthe whole simulation time for the decentralized andcentralized platoon networks in Fig. 5(d). A safetyviolation happens when there exist two consecutivevehicles in the platoon with the distance smaller than thesafety distance. We find that the total number of safetyviolations of the centralized platoon network is lagerthan that of the decentralized platoon network. With ahigher packet delay, vehicles are more likely to fail toadjust their velocities in time once their neighboringvehicles change their relative locations in the platoon,which causes more safety violations.

4.2. Channel allocation

We define packet delivery ratio as the percentage ofthe packets successfully delivered to their destinationvehicles in al the packets sent out during the wholesimulation time (i.e, 15 minutes). We define commu-nication cost as the total number of packets sent outduring the whole simulation time. In both the graph-based and SINR-based methods, the communication

cost includes 1) the packets that all the vehicles sendto the leader vehicle to inform their initial locations,2) the packets that the leader vehicle sends to notifyeach follower vehicle its allocated channels, and 3) thepackets that each vehicle sends to the leader vehiclewhen its location changes. In FLA, the communicationcost only refers to the number of packets that theleader vehicle sends to notify each vehicle the leadervehicle’s updated location. For each packet, we definethe packet delay as the time duration from the packetbeing sent to the packet being successfully delivered.And then, we calculate the average packet delay of allthe packets sent during the simulation. In this test, inevery minute, there is one vehicle entering the platoonand one vehicle leaving the platoon. Fig. 6(a),(b),(c),and (d) compare the packet delivery ratio (of each 10seconds), the communication cost, the average packetdelay, and the total number of safety violations ofthe three channel allocation methods with a differentnumber of vehicles. Comparing the four figures, wefind that FLA 1) produces the average packet deliveryratio almost the same as the SINR-based method buthigher than the graph-based method, and 2) generateslower communication cost, average packet delay and asmaller number of safety violations than both the graph-based and SINR-based methods. FLA has much lowercommunication cost because when the relative locationof a vehicle in the platoon changes so that its seg-ment changes, the vehicle can change its own channelbased on its scored FLA table without communicatingwith the leader vehicle. However, both graph-based andSINR-based methods require all vehicles to send tothe leader vehicle their locations. Also, when a vehiclechanges its relative location, it needs to send a notifi-

0.95

0.96

0.97

0.98

0.99

1

1.01

8 13 18 23 28P

ack

et

de

liv

ere

d

rati

oVelocity (m/s)

Graph-basedSINR-basedFLA

(a) Packet delivery ratio

0

2000

4000

6000

8000

8 13 18 23 28

Co

mm

un

ica

tio

n

co

st

# of vehicles

Graph-basedSINR-basedFLA

(b) Communication cost

0.00

0.50

1.00

1.50

2.00

8 13 18 23 28Av

e.

pa

ck

et

de

lay

(s

)

# of vehicles

Graph-basedSINR-basedFLA

(c) Average packet delay

0

50

100

150

200

250

8 13 18 23 28

To

tal

# o

f sa

fety

v

iola

tio

ns

# of vehicles

Graph-basedSINR-basedFLA

(d) Safety violation

Figure 6. Comparison of different channel allocation methods.

cation to the leader vehicle. Then, the leader vehiclerecalculates the channel allocation and sends the newchannel allocation to all the follower vehicles, whichsignificantly increases the communication cost. Morepackets generate more interference and hence lowerpacket delivery ratio. Furthermore, when more packetsare transmitted, each packet needs to wait longer timebefore other packets finish, leading to higher packetdelay. Vehicles are less likely to adjust their velocitiesin time with higher packet delay when their neighboringvehicles’ relative locations are changed in the platoon,leading to more safety violations. Hence, both graph-based method and SINR based method have higherpacket delay more safety violations than our method. Asa result, compared to the SINR-based method, FLA haslower communication cost, average packet delay andlager number of safety violations without compromisingthe packet delivery ratio, while compared to the graph-based method, our method has better performance inall aspects of communication cost, average packet de-lay, packet delivery ratio, and total number of safetyviolations.

5. Related Work

Vehicle networks. Several works proposed to improvethe performance of vehicle networks by controllingchannel congestion [27]–[29] or trying to decreasepacket drop rates [30]–[32]. All of these works considera centralized platoon network, where the leader vehiclecan support a limited number of vehicles into theplatoon and the number of vehicles inside the platoonis varied due to leader vehicle’s limited communicationcapability and its performance is crucial for vehiclesafety purpose. In contrast to the centralized platoonnetwork, proposed decentralized platoon network cansupport more vehicles with lower packet drop rate

and packet transmission delay, which provides bettervehicle safety.Channel allocation. Based on the choice of interfer-ence models, the current channel allocation methods canbe classified to two groups: graph-based scheduling andSINR-based scheduling. In the graph-based schedulingmethods [16], [24]–[26], for each pair of nodes, acentral node builds an edge between the two nodesiff they are within each other’s transmission range.Then, the central node partitions all the nodes intoseveral subsets, where the nodes in each subset arenot connected by any edge. Finally, the central nodeallocates a channel to each subset. These channel allo-cation methods are constrained by the limitations of thegraph interference model that ultimately abstracts awaythe accumulative nature of wireless signals. The SINRmodel offers a more realistic representation of wirelessnetworks. In the SINR-based scheduling method [13],[23], the central node allocates the same channel to theset of nodes such that, for each node, the accumulatedinterference from all other nodes is small enough forthis node to decode packet. However, all these SINR-based scheduling requires the location information ofall the senders in the network, which leads to highcommunication cost and long packet transmission delayif directly applied to the platoon network.

6. Conclusions

Centralized platoon networks cannot provide highvehicle safety or support a large number of vehiclesdue to the direct communication between the leadervehicle and each follower vehicle and the limited com-munication capacity of the leader vehicle. To overcomesuch problems, we proposed a decentralized platoonnetwork where each vehicle only needs to commu-

nicate with its two neighbor vehicles. However, thedecentralized platoon network brings the interferenceproblem: reducing the signal interference in simulta-neous multiple transmissions and maintaining the con-nectivity of platoons in vehicle dynamics. To handlethe interference problem, we designed the Fast andLightweight Autonomous channel selection algorithm(FLA), in which each vehicle determines its channelonly based on its distance from the leader vehicle. Oursimulation results show that the decentralized platoonnetwork can scale out well with low packet drop rate,low packet delay, and low safety violation. Also, FLAoutperforms the previous channel allocation methodsfor platoons in terms of packet delivery ratio, packetdelay, communication cost, and safety violation. In ourfuture work, we will study different channel allocationmodels for high-speed decentralized platoon network.

Acknowledgements

This research was supported in part by U.S. NSFgrants NSF-1404981, IIS-1354123, CNS-1254006, IB-M Faculty Award 5501145, and Microsoft ResearchFaculty Fellowship 8300751.

References

[1] B. Van Arem, C. J. Van Driel, and R. Visser, “The impact of co-operative adaptive cruise control on traffic-flow characteristics,”IEEE Trans. on ITS, vol. 7, no. 4, pp. 429–436, 2006.

[2] L. Yan and H. Shen, “Top: Vehicle trajectory based drivingspeed optimization strategy for travel time minimization androad congestion avoidance,” in Proc. of MASS, 2016.

[3] G. Guo and W. Yue, “Sampled-data cooperative adaptive cruisecontrol of vehicles with sensor failures,” IEEE Trans. on ITS,vol. 15, no. 6, pp. 2404–2418, 2014.

[4] S. Li, K. Li, R. Rajamani, and J. Wang, “Model predictivemulti-objective vehicular adaptive cruise control,” IEEE Trans.on CST, vol. 19, no. 3, pp. 556–566, 2011.

[5] C. Desjardins and B. Chaib-draa, “Cooperative adaptive cruisecontrol: A reinforcement learning approach,” IEEE Trans. onITS, vol. 12, no. 4, pp. 1248–1260, 2011.

[6] C. Sommer, S. Joerer, M. Segata, O. Tonguz, R. L. Cigno, andF. Dressler, “How shadowing hurts vehicular communicationsand how dynamic beaconing can help,” in Proc. of Infocom,2013, pp. 110–114.

[7] X. Yang, J. Liu, N. Vaidya, and F. Zhao, “A vehicle-to-vehiclecommunication protocol for cooperative collision warning,” inProc. of MobiQuitous, 2004, pp. 114–123.

[8] K. Yi and Y. D. Kwon, “Vehicle-to-vehicle distance and speedcontrol using an electronic-vacuum booster,” JSAE review,vol. 22, no. 4, pp. 403–412, 2001.

[9] E. Biglieri, J. Proakis, and S. Shamai, “Fading channels:Information-theoretic and communications aspects,” Tran. onInformation Theory, 1998.

[10] E. Ozatay, S. Onori, J. Wollaeger, U. Ozguner, G. Rizzoni,D. Filev, J. Michelini, and S. D. Cairano, “Cloud-based ve-locity profile optimization for everyday driving: A dynamic-programming-based solution,” IEEE Trans. on ITS, vol. 15,no. 6, 2014.

[11] C. Han, M. Dianati, R. Tafazolli, R. Kernchen, and X. Shen,“Analytical study of the ieee 802.11p MAC sublayer in vehicularnetworks,” IEEE Trans. on ITS, vol. 13, no. 2, pp. 873–886,2012.

[12] S. Hu, H. Liu, L. Su, H. Wang, T. F. Abdelzaher, P. Hui,W. Zheng, Z. Xie, and J. A. Stankovic, “Towards automaticphone-to-phone communication for vehicular networking appli-cations,” in Proc. of Infocom, 2014, pp. 1752–1760.

[13] O. Goussevskaia, Y. A. Osfwald, and R. Wattenhofer, “Com-plexity in geometric SINR,” in Proc. of Mobihoc, 2007.

[14] D. Chafekar, V. Anil Kumar, M. V. Marathe, S. Parthasarathy,and A. Srinivasan, “Capacity of wireless networks under SINRinterference constraints,” Wireless Networks, vol. 17, no. 7, pp.1605–1624, 2011.

[15] G. Pei and V. A. Kumar, “Low-complexity scheduling forwireless networks.” in Proc. of Mobihoc, 2012.

[16] A. Elbagoury, M. Youssef, D. Xue, and E. Ekici, “Demo v(t)csma: A link scheduling algorithm in wireless networks withimproved delay characteristics.” in Proc. of Mobicom, 2013.

[17] J. Liu, L. Yu, H. Shen, Y. He, and J. Hallstrom, “Character-izing data deliverability of greedy routing in wireless sensornetworks,” in Proc. of SECON, 2015.

[18] Z. Li and H. Shen, “Learning network graph of SIR epidemiccascades using minimal hitting set based approach,” in Proc. ofICCCN, 2016.

[19] S. M. Loos, A. Platzer, and L. Nistor, “Adaptive cruise control:Hybrid, distributed, and now formally verified,” in Proc. ofFormal Methods. Springer, 2011, pp. 42–56.

[20] “Matlab,” http://www.mathworks.com/products/matlab/, [Ac-cessed in March 2016].

[21] “Network simulator 3,” https://www.nsnam.org/ns-3-22/, [Ac-cessed in March 2016].

[22] A. Ghasemi, R. Kazemi, and S. Azadi, “Stable decentralizedcontrol of a platoon of vehicles with heterogeneous informationfeedback,” IEEE Trans. on VT, vol. 62, no. 9, pp. 4299–4308,2013.

[23] O. Goussevskaia, R. Wattenhofer, M. M. H. orsson, and E. Wel-zl., “Capacity of arbitrary wireless networks.” in Proc. of Info-com, 2009.

[24] C. Joo, X. Lin, and N. Shroff, “Understanding the capacityregion of the greedy maximal scheduling algorithm in multi-hop wireless networks,” in Proc. of Infocom, 2008.

[25] W. Wang, Y. Wang, X. Li, W. Song, and O. Frieder, “Efficientinterference aware TDMA link scheduling for static wirelessnetworks,” in Proc. of Mobicom, 2006.

[26] W. Cheng, X. Cheng, T. Znati, X. Lu, and Z. Lu, “JointTCP congestion control and CSMA scheduling without messagepassing,” in Proc. of Infocom, 2009.

[27] B. Scheuermann, M. Transier, C. Lochert, M. Mauve, andW. Effelsberg, “Backpressure multicast congestion control inmobile ad-hoc networks,” in Proc. of CoNEXT. ACM, 2007,p. 23.

[28] B. Scheuermann, W. Hu, and J. Crowcroft, “Near-optimal co-ordinated coding in wireless multihop networks,” in Proc. ofCoNEXT. ACM, 2007, p. 9.

[29] Y. Lin and H. Shen, “Vshare: A wireless social network aidedvehicle sharing system using hierarchical cloud architecture,” inProc. of IoTDI, 2016.

[30] C. Lei, E. van Eenennaam, W. K. Wolterink, G. Karagiannis,G. Heijenk, and J. Ploeg, “Impact of packet loss on CACC stringstability performance,” in Proc. of ITST, 2011, pp. 381–386.

[31] B. Kihei, J. A. Copeland, and Y. Chang, “Improved 5.9GHzV2V short range path loss model,” in Proc. of MASS. IEEE,2015, pp. 244–252.

[32] A. Sarker, C. Qiu, H. Shen, A. Gil, J. Taiber, M. Chowdhury,J. Martin, M. Devine, and A. Rindos, “An efficient wirelesspower transfer system to balance the state of charge of electricvehicles,” in Proc. of ICPP. IEEE, 2016.