Accepted Manuscript

A mobility-based scheme for dynamic clustering in vehicular ad-hoc networks (VANETs)

Mengying Ren, Lyes Khoukhi, Houda Labiod, Jun Zhang, Veronique Veque

PII: S2214-2096(16)30069-9
DOI: http://dx.doi.org/10.1016/j.vehcom.2016.12.003
Reference: VEHCOM 70

To appear in: Vehicular Communications

Received date: 1 June 2016

Revised date: 24 November 2016
Accepted date: 21 December 2016

Please cite this article in press as: M. Ren et al., A mobility-based scheme for dynamic clustering in vehicular ad-hoc networks (VANETs),
Veh. Commun. (2016), http://dx.doi.org/10.1016/j.vehcom.2016.12.003

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing
this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is
published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 A Mobility-based Scheme for Dynamic Clustering
in Vehicular Ad-hoc Networks (VANETs)
Mengying Ren∗ , Lyes Khoukhi∗ , Houda Labiod† , Jun Zhang† , and Veronique Veque‡
∗ University
of Technology of Troyes, France
{mengying.ren, lyes.khoukhi}@utt.fr
† Telecom ParisTech, University of Paris-Saclay, France
{mengying.ren, jun.zhang, houda.labiod}@telecom-paristech.fr
‡ University of Paris-Sud, France
veronique.veque@u-psud.fr

Abstract—Vehicle clustering is an efficient approach to improve safety applications aim to improve drivers and passengers
the scalability of networking protocols in vehicular ad-hoc comfort level and enhance traffic efficiency [2]. A detailed
networks (VANETs). However, some characteristics, like highly classification for road safety applications and their require-
dynamic topology and intermittent connections, may affect the
performance of the clustering. Establishing and maintaining ments are given in the standard of European Telecommunica-
stable clusters is becoming one of big challenging issues in tions Standard Institute (ETSI) [3].
VANETs. Recent years’ researches prove that mobility metric VANETs have several characteristics that distinguish them
based clustering schemes show better performance in improving from other multi-hop networks. Nodes in VANETs are highly
cluster stability. Mobility metrics, including moving direction, mobile, leading to a high probability of network partitions,
vehicle density, relative velocity and relative distance, etc., are
more suitable for VANETs instead of the received radio strength especially under highway scenarios. Therefore, the end-to-
(RSS) and identifier number metrics, which are applied for end communication cannot be guaranteed [4]. Intermittent
MANETs clustering. In this paper, a new dynamic mobility- connection may cause severe packet loss problem, and fur-
based and stability-based clustering scheme is introduced for ther influence traffic safety. Meanwhile, as a decentralized
urban city scenario. The proposed scheme applies vehicle’s self-organizing network, VANETs is lack of a centralized
moving direction, relative position and link lifetime estimation.
We compared the performance of our scheme with Lowest-ID and management and coordination entity which is responsible for
the most recent and the most cited clustering algorithm VMaSC managing the bandwidth and contention operations. Moreover,
in terms of cluster head duration, cluster member duration, VANETs is a large scale network; however, the communi-
number of clusters, cluster head change rate and number of cation range of a vehicle is limited which may also cause
state changes. The extensive simulation results showed that our a weak connectivity between nodes. Therefore, maintaining
proposed scheme shows a better stability performance.
a global network topology is indispensable for a node. For
Index Terms—Vehicular Ad Hoc Networks (VANETs), cluster- these reasons, a flat network topology is no longer effective
ing algorithm, wireless communication. for information transmission in VANETs [5]. To solve this
problem, a hierarchical network topology, called cluster, has
been proposed for VANETs.
I. I NTRODUCTION
A cluster is a virtual group of nodes having similar charac-
Vehicular ad hoc networks (VANETs), is a vital part of teristics. Clustering scheme is the method to divide vehicles
Intelligent Transportation System (ITS), which aims to im- into different groups according to some rules. Each cluster
prove road safety and information transmission efficiency on elects at least one leader, called cluster head, who serves as
the road. With the developments of automotive manufacturing, a local central management entity, performing intra-cluster
intelligent vehicle and wireless communication technologies, communication arrangement, local information aggregation,
vehicles which are equipped with wireless interfaces can and local information dissemination, etc. [4]. A cluster head
communicate with nearby vehicles directly through a V2V is followed by one or more than one cluster members. As a
(Vehicle-to-Vehicle) communication mode, as well as with hierarchical network, the first level of the network is called
fixed equipment, called Road Side Units (RSUs), through intra-cluster communication, where a cluster member can
a V2I (Vehicle-to-Infrastructure) or I2V (Infrastructure-to- directly communicate with its cluster head or nearby cluster
Vehicle) communication manner [1]. members within the same cluster. The second level of the
These types of wireless communications enable vehicles to network is inter-cluster communication, of which a cluster
share different kinds of information, including safety related head communicates with nearby cluster heads or road side
information and (non-safety related) infotainment information, infrastructures. Sometimes, cluster gateway node is proposed
corresponding respectively to road safety and non-safety appli- for neighboring cluster communication [6].
cations. Safety applications focus mainly avoiding accidents. In [6], a detailed survey of clustering schemes for VANETs
They require low latency and high reliability, whereas non- is well presented. Clustering scheme performance is usually
judged by cluster stability. Generally, clustering schemes, [9]. Lowest-ID [7] is a clustering algorithm originally pro-
providing high cluster stability, should ensure the following posed for MANETs. The CH is the vehicle which has the
properties: (1) lower transmission overhead; (2) longer cluster lowest identifier among its neighbors. Every node determines
head lifetime and longer cluster member lifetime; (3) less its cluster and only one cluster in a distributed manner.
average number of state changes per vehicle; (4) less cluster Lowest-ID provides single-hop and non-overlapping clusters
head changes. which reduce the use of bandwidth. One-hop cluster topology
In this paper, we present a simple dynamic mobility-based can reduce the cluster formation time and decrease cluster
clustering scheme in the purpose of establishing a stable management overhead while fewer information exchanges are
network backbone for future data aggregation and information required. However, the number of vehicles in a cluster usually
transmission. The proposed scheme is based on vehicles’ depends on vehicle’s transmission range and vehicle density.
mobility patterns, including moving direction, relative velocity, When vehicle density is very high, collision could happen
relative distance, and link lifetime. Different from previous in the cluster and would cause high packet loss rate. When
clustering schemes of which nodes are static during cluster vehicle density is very low, a vehicle is unable to detect
formation, our scheme proposes a dynamic cluster formation neighbors. To solve this problem, VWCA [10] proposed an
procedure. A ”temporary cluster head” is proposed to help adaptive transmission range algorithm (AART), based on the
cluster formation. In addition, we propose a ”safe distance intra-cluster communication standard, Dedicated Short-range
threshold” in order to control the cluster size. The proposed Communication (DSRC) standard [11]. Vehicle can change
clustering scheme is evaluated in terms of cluster stability, and the transmission range dynamically from 100m to 1000m
its performance is compared with Lowest-ID [7] algorithm. according to vehicle density.
The rest of the paper is organized as follows: Section In recent years, researchers focus on building up multi-hop
II discusses the related work in VANET clustering. Section clusters. In [12], Zhang et al. firstly proposed a multi-hop
III presents the new proposed clustering algorithm from the clustering algorithm for VANETs, based on vehicle’s relative
aspects of cluster head selection, cluster formation and cluster mobility. The cluster size is limited by the number of hops
maintenance. Section IV presents the simulation environment between CH and its farthest CM. The algorithm proposed in
and the performance analysis of our scheme. Section V [13], known as VMaSC, is addressed as the first multi-hop
concludes the paper and briefly introduces our future work. clustering scheme to simulate under realistic traffic scenario,
which is generated by Simulation of Urban MObility (SUMO)
[14]. The scheme aims to provide more stable clusters and to
II. RELATED WORK
reduce the number of CHs in the network. The CH election
Generally, clustering procedure can be separated into cluster is based on the calculated relative mobility with respect to its
head (CH) selection, cluster formation, and cluster mainte- neighbors. The performance of VMaSC was compared with
nance. CH selection allows vehicles to choose the CH in a [12] for 1-hop, 2-hops, 3-hops, and VMaSC shows a better
centralized or distributed manner; the selection criteria are performance in terms of CH duration, CM duration and CH
presented in the following part of this section. Cluster forma- change times, especially when the cluster size is set to 3-hop.
tion process aims to establish the communication link between Generally, compared to single-hop clustering schemes, a multi-
CH and its cluster members (CMs). Normally, a stable cluster hop clustering scheme requires more Beacon exchanges within
requires a stable link between CH and CMs. However, because the maximum number of hops, which may cause the increase
of the dynamic nature of VANETs, individual links may come in the number of connections lost and longer cluster formation
into existence and vanish unpredictably, making the task of time. Simulation results in VMaSC [13] show that the cluster
establishing and maintaining communication between fast- stability decreases considerably when the maximum number
moving vehicles very challenging [8]. Cluster maintenance of hops is above 3. [15] proposed a neighborhood following
process focuses on solving the cluster re-formation and vehicle strategy for multi-hop clustering, in which, each vehicle finds
re-affiliation problems. A good cluster maintenance scheme a stable target to follow. The vehicle only needs to know the
should generate less control overhead and should not use information of its local one-hop neighbors, thus, reduces the
too much network resources. This section introduces some packet loss problem.
clustering algorithms in VANETs.
B. Clustering metrics
A. Cluster topology Clustering schemes can also be classified based on clus-
In terms of cluster topology, clustering schemes can be tering metrics, including CH selection metrics and cluster
classified into single-hop clustering and multi-hop clustering, formation metrics. A simple and direct way to choose a CH is
indicating the maximum number of hops from a cluster head selecting the first vehicle moving in a certain direction. Cluster
to its farthest cluster member. platooning in CONVOY [16], proposed for highway scenarios
The majority of clustering algorithms are based on single- selects the first vehicle as a CH. Vehicles within the prede-
hop clusters in which the CM is one-hop away from its CH. fined maximum distance to CH are combined together, which
CH only chooses its CMs from the local vicinities. Vehicle construct a multi-hop cluster. MC-DRIVE [17] proposed a
information are broadcasted through periodic beacon messages direction-based clustering algorithm for intersection area. The
first vehicle moving in a certain direction was selected as CH; and average state change rate per node, etc.. Cluster stability is
clusters are formed in one-hop based on CHs’ transmission a crucial measure of the efficiency of clustering algorithms for
range (TR). However, this simple CH selection mechanism is VANETs. In [26], the author presented a stochastic analysis
only suitable for simple road topology, like straight highway. of the vehicle mobility impact on single-hop cluster stability,
Instead of simply choosing the first vehicle as CH, most and a stochastic mobility model was proposed.
clustering mechanisms prefer calculating the stability of a In this paper, a new mobility-based clustering algorithm
node to its surroundings. MOBIC [18] was the first article for VANETs is presented. We evaluate cluster stability from
proposing aggregate mobility (It was originally proposed for the following aspects: average cluster head lifetime, average
ad hoc networks). Each node calculates its relative mobility to cluster member lifetime, average state change rate per node,
all of its neighbors based on Received Signal Strength (RSS). cluster head change rate, and average number of clusters.
The node with the lowest aggregate mobility is chosen as the
CH. Similar to MOBIC, the New-ALM [19] also chooses a III. P ROPOSED APPROACH
node with less variance relative to its surroundings as a CH. The paper focuses on proposing a new clustering algorithm
Instead of using the RSS parameter, New-ALM calculated based on V2V communication for urban city scenario. It
relative distance between nodes. Later, to improve the cluster assumes that every vehicle is equipped with an On Board Unit
stability, the paper [12] proposed a K-hop clustering. K-hop (OBU) wireless transceiver/receiver and has a GPS receiver
relative mobility is based on the ratio of packet delivery that can update vehicle’s location on the road. Meanwhile,
delay of two consecutive packets. PPC [20] is also a multi- each vehicle can calculate the relative velocity with respect to
hop clustering mechanism which is based on vehicles’ speed its one hop neighbors, as well as detect the relative distance
variations and the predicted traveling time. Vehicle’s ”Eligibil- to its vicinities.
ity” value, indicating cluster stability, decreases exponentially
with the increased speed deviation. APROVE [21] is based A. Cluster definition
on a data clustering technique, called Affinity Propagation We suppose that vehicles enter the road segment one by one
(AP) [22]. Each vehicle sends hello messages periodically, with a predefined traffic flow rate. Each vehicle moving on the
aggregating availability and responsibility messages. Vehicles’ road broadcasts a Beacon message at every Beacon Interval
relative distance, position and prediction position of near (BI). According to the clustering metrics we have mentioned
future are used in [21]. Vehicle with highest sum of availability in Section II-B, cluster head (CH) should be the vehicle
and responsibility value is selected as a CH. Moreover, a which has higher relatively stability among its neighboring
cluster contention time (CCI) is proposed when two CHs vehicles. Therefore, we choose the vehicle nearest to the
encounter each other in order to reduce the unnecessary cluster central geographical position of a cluster as the CH, so that
reformation. SCRP [23] is a cluster-based routing protocol its neighbors should spend more travel time to leave the
using Dominating Set (DS), which attempts to select a small cluster, and the cluster is considered to be more stable. Cluster
number of mobile nodes as dominating nodes to form a stable members (CMs) are selected from CH’s one-hop neighbor set.
backbone in a network.
Some other clustering mechanisms are based on Weighted
Clustering Algorithm (WCA). The CH selection is based on
the weighted sum operation. In [24], the author proposed a
lane-based clustering algorithm based on vehicles’ relative
speed, relative position and traffic flow. Each lane can be
distributed with a certain weight according to the traffic flow.
VWCA [10] calculates the weighted clustering value based on
the metrics: vehicle distrust value, entropy value, number of
neighbors and relative position. The vehicle with the minimum
weighted sum value in the neighbor is selected as CH. Another
weighted clustering mechanism AMACAD [25] was proposed Fig. 1. Clusters (T R: Transmission Range; L: cluster length; Dt : Safe
based on vehicle’s final destination. In AMACAD, vehicles Distance threshold; GW i, GW b: Gateway node.)
with similar destinations have higher possibility to stay in Fig. 1 shows two clusters on a straight road, cluster Ci and
the same cluster. The weighted sum is calculated based on cluster Ci+1 (clusters are represented by rectangles). Cluster
vehicles’ relative destinations, final destinations, relative speed head is in the central position, and the length of the cluster
and current position. is smaller than twice of CH’s transmission range (TR). In
our proposed clustering scheme, each cluster consists of two
C. Cluster performance evaluation gateway nodes moving on the edge of the cluster: one is
As well as we know, the majority of clustering algorithms moving ahead and another one is moving in the end of the
proposed for VANETs focus on improving the cluster stability. cluster.
Generally, cluster stability can be evaluated from the following Due to the rapid changes of vehicle mobility, vehicles on
aspects: CH duration, CH change rate, re-clustering frequency, the edge of CH’s transmission range are considered not being
stable enough, and may cause frequent CM disconnections and until TU N expires, it changes its state to CHt; otherwise, it
CM re-clustering. To solve this problem, we introduce a ”Safe changes the state to CM upon receiving a confirmation beacon
Distance Threshold”, denoted as Dt , which should be smaller message, called Beacon ACK message, from a CH or a CHt.
than vehicle’s TR, Dt ≤ T R. Therefore, the vehicles within
Dt range of the CH are considered having more stable links
with their CH. The size of the cluster is defined as L ≤ 2Dt .
Table I lists the notations used through this study.

TABLE I
N OTATIONS

Notation Description
TR Transmission range
BI Beacon Interval
Dt Safe Distance threshold
MI Merge Interval
Vi Vehicle i
Ci Cluster i
CHi Cluster head i
CMi Cluster member i
CHti Temporary cluster head i
U Ni Undecided Node i
Dir(i) Moving direction of Vi Fig. 2. State transition machine
ΔDij Relative distance between Vi and Vj
Li Length of cluster Ci The CHt vehicle sets a timer TCHt and initiates a cluster
TU N Timer for UN transfer to CHt
TCHt Timer for CHt transfer to CH formation process which will be described in the next section.
Twb Timer for CH to monitor Beacon from its CM Upon Beacon ACK message reception from a CH, CHt will
CID Cluster ID change its state to CM if it does not have any followers,
CM Li CM list of CHi
Lmerge Length of the merged cluster CM L = ∅. In another situation, the CHt vehicle changes to
BLi Beacon list of CHi , recording the received Beacons CM during a CH selection procedure, described in the next
section. Otherwise, the CHt vehicle changes to CH when TCHt
B. Cluster state transition expires.
When a CHi hears a Beacon ACK message from a
In the proposed clustering algorithm, a vehicle may have
neighboring CHj , it checks whether it has CMs or not. If
one of the following 4 states: Undecided Node (UN), Cluster
CM Li = ∅, it changes its state to a CM of CHj . Furthermore,
Head (CH), Cluster Member (CM), and Temporary Cluster
when cluster merging happens, a CH vehicle can also change
Head (CHt). The states are specified in the following:
state to a CM of the merged cluster.
• UN: Initial state of all vehicles, which means that the
The CM vehicle will change the state to CH when it
vehicle does not belong to any clusters.
receives CH notif ication message from its CHt, or when it
• CH: The leader of the cluster, which can communicate
is selected as CH in the merged cluster during cluster merging
with all of its members. Each cluster has only one CH
process. A CM vehicle can hear Beacon ACK messages
and each CH maintains a CM list, CML, recording the
periodically from its CH or CHt, otherwise, it changes the
information of its CMs.
state back to UN if it is no longer the CM of the current
• CM: The normal vehicle which is a one-hop neighbor of
cluster, which will be described in Section III-F2.
a CH. A particular type of CM is the gateway node (GW),
which is responsible for inter-cluster communication and
C. Cluster formation
is located on the edge of the cluster. Each cluster may
have two gateway nodes: GWi , moving ahead of the As been shown in Fig. 3, a vehicle i in the UN state,
cluster, and GWb , moving in the end of the cluster. U Ni , tries to join an existing cluster by listening to the
• CHt: The temporary CH vehicle. It only appears at the Beacon message from a CH or CHt during the time period
beginning of cluster formation procedure and disappears TU N . If U Ni fails to join an existing cluster when TU N
when the CH is elected. expires, it claims itself as a CHt node CHti , and sets its CID.
Fig.2 illustrates the possible state transitions of a vehicle. Meanwhile, CHti starts a timer TCHt and begins a cluster
The vehicle starts with an UN state and sets a timer TU N , formation procedure, as described in Section III-D.
during which it hears Beacon message from a CH or a CHt. During the time period TU N , if U Ni hears a Beacon ACK
In our study, Beacon messages, which are broadcasted by message from CHj or CHtj , it checks whether it is on the
CH or CHt vehicles, are denoted as Beacon ACK message, ACK list. If yes, U Ni changes its state to CM directly
aggregating confirmation information. Each Beacon ACK and sets its cluster identifier CID = j; otherwise, it checks
message contains an ACK list, a list of node identifiers. If whether it is a CM candidate of CHj or CHtj . In this paper,
the UN vehicle does not hear any Beacon ACK message the CM candidate should be the vehicle which are moving in
 Algorithm 1 CH selection process
while TCHt = 0 && CHtj is still in state CHt do
if CHtj receives ReqJoin from U Ni then
if U Ni is moving behind CHtj && ΔDij <= Dt
then
CM Lj ← U Ni
ACK list ← U Ni
CHtj broadcasts Beacon ACK at next BI
else
if U Ni is moving behind CHtj && ΔDij > Dt
&& CM Lj = ∅ then
CHtj chooses the farthest vehicle CMk from
CM Lj
CHtj sends CH notif ication to CMk
CHtj → CMj
CID ← k
end if
end if
end if
end while
if TCHt == 0 && CHtj is still in state CHt && CM Lj =
∅ then
CHtj → CHj
Fig. 3. Cluster formation
end if
the same direction with its CH, Dir(i) = Dir(j). U Ni sends
When CMk receives CH notif ication
a ReqJoin message to CHj or CHtj , if it is a CM candidate.
Upon receiving a ReqJoin message from vehicle Vi (Vi CMk → CHk
could be in the state UN or CH), CHj checks the following CM Lk ← CM Lj
conditions to confirm that the requester is a qualified CM: (1) CID ← k
the relative distance between Vi and CHj , ΔDij , should be CHk broadcasts Beacon ACK
smaller than the predefined Dt , ΔDij ≤ Dt ; (2) Vi is not on
CM Lj (the CM list of CHj ).
If Vi is a qualified CM, CHj adds the information of Vi E. Gateway node selection
into its CM list CM Lj . Meanwhile, CHj adds the identifier As soon as the CH is selected and the cluster is well formed,
of Vi to its confirmation list, ACK list, which will be CHk selects two CMs, which are moving on the edge, to
broadcasted within its next Beacon ACK message. It is be the GW nodes. However, it happens sometimes that two
noticed, only vehicles in the state CH or CHt can broadcast a GW candidates have the same relative distance from their CH.
Beacon ACK message. To solve this problem, we introduce an estimated connection
time between CH and CM, called link lifetime (LLT), to
D. Cluster head selection
evaluate the link sustainability. A higher LLT represents a
Similar to a CH, CHt can also add qualified CMs according more sustainable link. CH will select the GW node which
to the conditions mentioned above. However, CHt only adds has larger LLT value. The work in [27] gives the definition of
CMs which are in its neighborhood. Algorithm 1 describes a LLT, shown in Eq. (1), when two vehicles are moving in the
CH selection procedure initiated by a vehicle CHtj . same or opposite directions. Although vehicle position should
Upon receiving a ReqJoin message from U Ni moving be represented by x-coordinate and y-coordinate, this study
behind, if CHtj detects the relative distance ΔDij > Dt assumes the trajectory of all vehicular nodes to be a straight
and CM Lj = ∅, CHtj selects the farthest CM CMk in its line, as the lane width is small. Thus, the y-coordinate can be
CM Lj to be the CH, and sends a CH notif ication message, ignored. We denote the positions of Vk and Vj by xk and xj ,
containing its CM list CM Lj , to inform CMk to become respectively.
CHk . Meanwhile, CHtj changes to CMj state and resets
CID = k. After receiving a CH notif ication message, −Δvkj ∗ ΔDkj + Δvkj ∗ T R
LLTkj = 2 (1)
CHk adds CMs and broadcasts a Beacon ACK to inform its (Δvkj )
CMs to reset CID = k. CHk continues the cluster formation
process via adding new CMs. In another case, if CHtj is still ΔDkj = |xk − xj | (2)
in the state CHt and CM Lj = ∅ when TCHt expires, CHtj
claims itself as CHj . Δvkj = |vk − vj | (3)
 Note that the T R is the transmission range of the vehicle, Algorithm 2 Cluster merging process
vk and vj are the velocities of CHk and CMj , respectively. Upon receiving ReqM erge
CHi estimates Lmerge
F. Cluster maintenance if Lmerge ≤ 2 ∗ Dt then
Due to the high dynamic nature of VANETs, vehicles keep CHi selects central CM CMm to be CHmerge
joining and leaving clusters frequently, thus, causing extra CHi and CHi+1 sends ACK merge along with their
maintenance overhead. In our proposed scheme, Clusters are CM L to CMm respectively
dynamically moving on the road, with their CH inside of CHi → CMi
the clusters. When CH loses all of its CMs, it becomes an CHi+1 → CMi+1
UN node. Otherwise, it remains as CH until cluster merging CID ← m
process happens. Therefore, in our proposed scheme, the end if
cluster maintenance procedure only deals with cluster merging
and vehicle leaving steps. Upon receiving ACK merge
1) Cluster merging: The proposed algorithm allows cluster CMm → CHmerge
to be overlapped. However, when two neighboring clusters Ci CM Lm ← CM Li , CM Lm ← CM Li+1
and Ci+1 have a big overlapping area, as presented in Fig. CID ← m
4, cluster merging procedure is triggered. Instead of having CHmerge broadcasts Beacon ACK
two CHs, a single CH is selected. When the distance of two
CHs is smaller than the predetermined threshold Dt , cluster
merging procedure begins. To avoid frequent re-clustering, beacon list (BL) in order to record the reception of its CMs’
cluster merging is deferred. Instead of starting the cluster Beacons. Once a CH, for example CHi , receives a Beacon
merging procedure immediately, the merging procedure begins message from CMj , it checks whether CMj is within the
if two CHs can always hear each other and are always within range of Dt . If ΔDij ≤ Dt , CHi updates the information
the range of Dt during the Merge Interval (MI). Once the of CMj and set BLi (j) to 1, indicating the reception of the
cluster merging process begins, CHi+1 , moving behind, will information of CMj ; otherwise, it deletes CMj from CM Li .
send a ReqM erge message to CHi , the CH moving ahead.
Cluster merging process is described in Algorithm 2. Algorithm 3 Leaving a cluster
When CHi receives Beacon from CMj
if CMj ∈ CM Li && ΔDij ≤ Dt then
BLi (j) ← 1
CHi updates the information of CMj in CM Li
end if

Every time when Twb == 0

for all CMk ∈ CM Li do
if BLi (k) == 0 then
CHi deletes CMk and BLi (k)
Fig. 4. Cluster merging (Lmerge : length of the merged cluster; Dt : Safe else
Distance threshold.) if BLi (k) == 1 then
Upon ReqM erge message reception from CHi+1 , CHi es- BLi (k) ← 0
timates the potential merged cluster size Lmerge . If Lmerge ≤ end if
2Dt , cluster merging is permitted and a CH for the merged end if
cluster, called CHmerge , is selected, which is the nearest end for
node to the geographical central position of the merged Restart Twb
cluster. After selecting CHmerge , previous CHs will send a
ACK merge message, containing their CMs list CM L, to 3) CML and GW updating: Every time when a CH receives
CHmerge , and claims themselves as the CMs of CHmerge . a Beacon message from its CM, it updates CM’s information,
The CHmerge adds all of the CMs to its CM L and broadcasts for example the position, in its CM L. Therefore, every
a Beacon ACK message to inform its CMs to change their CH can monitor its CM L dynamically. Once the CM L is
CID. updated, GWi and GWb selection functions are triggered
2) Leaving a cluster: In the proposed approach, each CH immediately and cluster’s gateway information will also be
creates and updates a CM L dynamically. CH has to monitor updated according to the process described in Section III-E.
the presence of its CMs per every waiting beacon interval,
denoted as Twb . Therefore, CH can detect CM disconnection G. Important messages
as long as it does not receive the Beacon message from its Table II presents a set of important messages transmitted
CM at least Twb time period. Moreover, each CH creates a during the clustering procedure, and the message dissemi-
nation types are demonstrated. Every message must contain For each testing scenario, simulation runs for 800s. The
the following parameters: message type, source ID, source clustering process starts at time Tstart , the time when all
state, cluster identifier CID, x-coordination x, y-coordination vehicles have entered the road, and ends at time Tend , before
y, velocity v, and direction Dir. Compared to a simple which most of vehicles are still on the road. According to the
Beacon message, Beacon ACK adds a ACK list, and is testing scenarios, we set Tstart = 160s and Tend = 460s.
only broadcasted by a CH or CHt. Therefore, the clustering simulation time is 300s. All of
the simulations run 10 times. According to previous related
TABLE II works(e.g., [12], [13], [21]), our simulation parameters are
L IST OF IMPORTANT MESSAGES selected as illustrated in Table III.
Name of the message Source Dissemination type
Beacon UN or CM Broadcast TABLE III
Beacon ACK CH or CHt Broadcast S IMULATION PARAMETERS
ReqJoin Any single node Towards a CH or CHt
ReqM erge CH Towards a CH Parameter Value
CH notif ication CH or CHt Towards a new merged CH Simulation time 300 s
ACK U N CH Towards a CM MAC protocol IEEE 802.11p
TR 200 m
Number of vehicles 100
Road length 15 km
Length of car 5m
IV. S IMULATION Acceleration rate 2.6 m/s2
Deceleration rate 4.5 m/s2
In this section, we provide a deep analysis of our proposed Maximum lane speed (MLS) 10-40 m/s
clustering scheme, and compare the clustering performance Traffic flow rate (TFR) 1200 vehicles/hour
Dt 100-200 m
with two existing algorithms, Lowest-ID [7] and VMaSC [13]. BI 1.0 s
Since both the proposed algorithms and Lowest-ID are based MI 10.0 s
on one-hop cluster, the one-hop VMaSC is implemented in our Twb 5.0 s
Propagation model Two-Ray Ground
simulation. All of the clustering algorithms are implemented Number of iterations 10
on NS2 [28], and the testing scenarios are all generated by Mobility model Car-following model
Simulation of Urban MObility (SUMO) [14].
In the testing scenarios, the road topology consists of a two-
A. Performance metrics
lane and two-way road of length 15 km. Vehicles are deployed
in the road with a predefined traffic flow rate (vehicles per The clustering performance metrics, used for cluster stabil-
hour), denoted as TFR. The maximum vehicle velocity, being ity evaluation and comparison, are described as follows:
allowed on the road, is called maximum lane speed (MLS). • Average number of clusters: as long as the CH is alive,
We consider 100 vehicles, 50 vehicles for each direction. there is a cluster. This metric allows us evaluating the
We firstly evaluate the impacts of ”Safe Distance threshold” quality of cluster formation. In the worst case, each vehi-
Dt . Traffic flow rate is set to 1200 vehicles per hour, and cle represents an independent cluster; therefore, clustering
maximum lane speed is set to 20 m/s, which is considered a algorithm is meaningless.
regular speed on the road. The value of Dt is set to be in the • Average CH duration: this metric represents the cluster’s
range of 100-200m, smaller than vehicle’s transmission range. lifetime, the time interval between a vehicle becoming a
Therefore, cluster size is in the range of 200-400m, as defined CH and changing to another state. In general, a longer
in our algorithm. duration of CH represents a more stable cluster. In
The second simulation evaluates the impacts of the Bea- this paper, the normalized average CH duration is the
con Interval (BI) on the cluster stability with the increased percentage time period of the total simulation time.
maximum lane speed (MLS). The set of MLS are specified as • Average CM duration: it defines the average time interval
follows: 10, 15, 20, 25, 30, 35, 40 m/s; traffic flow rate is set from a node joining an existing cluster as a CM to
to 1200 vehicles per hour. The BI is set to 0.5s, 1.0s and 2.0s leaving the connected cluster or to becoming a CH. The
respectively. normalized average CM duration is the percentage time
In the third simulation, we evaluate the impacts of the period of the total simulation time.
maximum lane speed (MLS) on cluster stability, and compare • Average CH change rate (per second): the cluster head
the clustering performance of the proposed algorithm with change rate defines the number of state transitions from
Lowest-ID [7], denoted as LID, and one-hop VMaSC [13], CH to another state per unit time.
denoted as VMaSC 1hop under the same context. The set of • Average state change (per node): this metric indicates the
maximum lane speed are specified as follows: 10, 15, 20, 25, number of state transitions in each vehicle during the
30, 35, 40 m/s. The traffic flow rate is set to 1200 vehicles per clustering procedure.
hour. To make a fair comparison, Twb is set to 5.0 s, the same • Clustering efficiency: it is defined as the percentage of
value as CH T IM ER when implementing VMaSC [13], and vehicles participating in clustering procedure (vehicles
the same value of information updating interval in LID [7]. which are not in UN state) during the simulation. A
 higher clustering efficiency means a better clustering 100
CHnbr BI=0.5s
CHnbr BI=1.0s
performance.

Average number of
80 CHnbr BI=2.0s
CMnbr BI=0.5s
• CM disconnection frequency (per second): it illustrates

vehicle state
CMnbr BI=1.0s
60 CMnbr BI=2.0s

the total number of times that CMs lose the connections CHtnbr BI=0.5s
CHtnbr BI=1.0s
40 CHtnbr BI=2.0s
to their current CHs per unit time.
20

B. Results and analysis 0

10 15 20 25 30 35 40
Maximum Lane Speed (m/s)

100 (a) Average number of vehicles in each state

CHnbr CMnbr CHtnbr UNnbr
Number of vehicle state

80
1 1

Average CM duration (%)

Average CH duration (%)
60
0.8 0.8

40
0.6 0.6

20 0.4 0.4

0 BI=0.5s BI=0.5s
200 240 280 320 360 400 0.2 0.2
BI=1.0s BI=1.0s
Cluster size 2*Dt (m) BI=2.0s BI=2.0s
0 0
10 15 20 25 30 35 40 10 15 20 25 30 35 40
(a) Average number of vehicles in each Maximum Lane Speed (m/s) Maximum Lane Speed (m/s)
state
(b) Average CH duration (%) (c) Average CM duration (%)
1 1
Fig. 6. Impact of BI on cluster performance with the increased maximum
Average CM duration (%)
Average CH duration (%)

0.8 0.8
lane speed (MLS)
0.6 0.6

0.4 0.4
the number of CM vehicles decreases. This is because that
0.2
MLS=20 m/s
0.2
MLS=20 m/s
with the increased vehicle velocity, some CMs may move
TFR=1200 v/h TFR=1200 v/h
0
200 240 280 320 360 400
0
200 240 280 320 360 400
out of the cluster and may become isolated vehicles. Then, if
Cluster size 2*Dt (m) Cluster size 2*Dt (m)
the isolated vehicle cannot successfully re-connect to another
(b) Average CH duration (%) (c) Average CM duration (%) existing cluster, a new cluster will be formed, increasing the
Fig. 5. Impacts of Dt on cluster performance
number of CHs.

1) ”Safe Distance threshold” Dt: Fig. 5 presents the im- 1

CH
Average role duration (%)

pacts of ”Safe Distance threshold” Dt . In Fig. 5(a), with the 0.8

CM
CHt
increased Dt , less clusters are organized during the simulation. 0.6
This is because that more vehicles are combined in a cluster
0.4
as CMs when the cluster length increases under the same
traffic density. The numbers of vehicles in CHt and UN states 0.2

(both are temporary states) remain stable when cluster size 0

10 15 20 25 30 35 40
Maximum Lane Speed (m/s)
is becoming larger. Fig. 5(b) shows the average CH duration,
represented as the percentage of total simulation time. The Fig. 7. Average duration of each vehicle state of the proposed scheme
average CH duration increases slightly but remains relatively
stable, when Dt increases. Fig. 5(c) illustrates that the average
CM duration decreases slightly with the increased cluster size. 1 1
Our scheme
Average CM duration (%)

Our scheme
Average CH duration (%)

LID
We observe that Dt has small impacts on both the CH duration
LID
0.8 0.8 VMaSC_1hop
VMaSC_1hop

and CM duration. 0.6 0.6

2) Beacon Interval (BI): According to ETSI standard [29], 0.4 0.4

the Cooperative Awareness Message (CAM) is broadcasted 0.2 0.2

with the frequency 1-10Hz (0.1s-1s). Therefore, in our simu- 0

10 15 20 25 30 35 40
0
10 15 20 25 30 35 40
Maximum Lane Speed (m/s)
lation, we set BI to 1.0s as the default value, and change BI Maximum Lane Speed (m/s)

to 0.5s and 2.0s respectively, in order to evaluate its impacts (a) Average CH duration (%) (b) Average CM duration (%)
on our proposed algorithm.
Fig. 8. Vehicle state lifetime comparison under the impact of vehicle’s MLS
The results in Fig. 6 show the cluster performance in terms
of average number of vehicles in each state (Fig. 6(a)), average 3) Impact of maximum lane speed: Fig. 7 presents the
CH duration (Fig. 6(b)), and average CM duration (Fig. 6(c)). averaged lifetime of each vehicle state with the increased
From the results, we observe that BI has slight impact on the maximum lane speed (MLS), in the proposed algorithm. We
cluster performance. According to the simulation results, BI observe that when vehicle velocity increases from 10m/s to
is set to 1.0s in the rest of the simulation. In Fig. 6(a), we 40m/s, vehicle state lifetime is relatively stable. The CHt
observe that the number of vehicles in the state CH increases lifetime is very small because it is a temporary state which
with the increased maximum vehicle velocity, and meanwhile, only appears in the beginning of a cluster formation process.
 60 100

Average number of clusters

Our scheme 50 Our scheme

Average number of UNs

Average number of CMs
50 LID LID
VMaSC_1hop 80 VMaSC_1hop
40
40
60
30
30
40 20
20
Our scheme
10 20 10
LID
VMaSC_1hop
0 0 0
10 15 20 25 30 35 40 10 15 20 25 30 35 40 10 15 20 25 30 35 40
Maximum Lane Speed (m/s) Maximum Lane Speed (m/s) Maximum Lane Speed (m/s)

(a) Average cluster number (b) Average number of CMs (c) Average number fo UNs

Fig. 9. Vehicle state number comparison under the impact of vehicle’s MLS

CM disconnection frequency (/s)

1 5 0.8 1.2
Number of state changes

Our scheme

Clustering efficiency (%)

Our scheme
LID 0.7 1
VMaSC_1hop
CH change rate (/s)

0.8 4
VMaSC_1hop 0.6
(per vehicle)

0.8
0.6 3 0.5

0.4 0.6
0.4 2
0.3 0.4
Our scheme 0.2 Our scheme
0.2 1
LID 0.2 LID
VMaSC_1hop 0.1
VMaSC_1hop
0 0 0 0
10 15 20 25 30 35 40 10 15 20 25 30 35 40 10 15 20 25 30 35 40 10 15 20 25 30 35 40
Maximum Lane Speed (m/s) Maximum Lane Speed (m/s) Maximum Lane Speed (m/s) Maximum Lane Speed (m/s)

(a) Average CH change rate (per sec- (b) Average state change (per node) (c) CM disconnection frequency (per (d) Clustering efficiency (%)
ond) second)

Fig. 10. Cluster stability comparison under the impact of vehicle’s MLS

The results in Fig. 8, 9, and 10, compare the cluster stability Fig. 10 demonstrates the details of state transitions during
among the proposed clustering algorithm to Lowest-ID (LID) the clustering process. The results in Fig. 10(a) and Fig. 10(b)
and VMaSC (VMaSC 1 hop), from the aspects of the vehicle reveal that both the CH change rate and vehicle state change
state duration, the number of vehicle states, and the number times of LID grow quickly when MLS is increasing. In LID,
of state changing, respectively. CH changes its state as soon as it detects a neighbor vehicle
Fig. 8(a) and Fig. 8(b) reveal the impacts of the maximum having an identifier smaller than itself. When vehicle veloc-
vehicle velocity on the averaged CH lifetime and CM lifetime. ity increases, vehicle’s neighbor list changes considerably,
It is obvious that both the averaged CH and CM duration of causing more frequent CH change rate. CH change rates of
LID and VMaSC 1hop decrease rapidly when MLS increases. VMaSC 1hop and our scheme are both very low and remain
In Fig. 8(a), when vehicles move slowly on the road, both relatively stable in Fig. 10(a). In our scheme, CH may change
the mean CH duration of LID and VMaSC 1hop are higher to a CM when cluster merging happens or when it loses all
than that of our scheme. However, their CH duration decreases of its CMs, as mentioned in Section III. In Fig. 10(b), the
rapidly with the increased MLS, especially for LID. The CH number of state transitions for each vehicle in VMaSC 1hop
duration of VMaSC 1hop is always higher than that of our and our scheme are higher than that of LID. This is because
scheme until MLS becomes bigger than 33 m/s. This is more vehicle states are defined in these two schemes compared
because in our scheme, CHt assists cluster formation and to LID.
CH is selected during the cluster formation process, while The CM disconnection frequency, shown in Fig. 10(c),
CH selection happens in the beginning in VMaSC 1hop and presents a similar growth trend compared to the results in
LID. In Fig. 8(b), the CM duration in our scheme remains Fig. 10(b). It is because that vehicle state transition always
the highest one when MLS is bigger than 17 m/s. When happens when a CM loses the link connection with its current
MLS becomes bigger than 30 m/s, the CM duration of our CH. Since state transition in LID is identifier-based, Fig.
scheme is almost two times of VMaSC 1hop. The results in 10(c) only compares our scheme and VMaSC 1hop. It is
Fig. 9(b) and Fig. 9(c) well explain this consequence that obvious that our scheme shows a very low CM disconnection
with the increased MLS, cluster becomes less stable and many frequency compared to VMaSC 1hop, indicating that our
vehicles change to the temporary state in VMaSC 1hop [13], scheme provides higher cluster stability.
therefore, reduces the average CH and CM duraiton. From the results in Fig. 10(d), we observe that both LID and
Results in Fig. 9 show that both the number of the CHs our scheme perform a very high clustering efficiency, which
and the number of UNs in our scheme are slightly lower than is close to 100% when MLS increases. It means that almost
the results of LID. Moreover, when MLS becomes larger, all of vehicles on the road participate in clustering procedure
many CHs and CMs in VMaSC 1hop change to UN state during the simulation. However, with the growth of MLS, the
(SE state in [13]). Therefore, the number of CHs and CMs of clustering efficiency of VMaSC 1hop decreases significantly.
VMaSC 1hop in Fig. 9(a) and in Fig. 9(b) decrease and the It is because the number of UN nodes increases rapidly when
number of UNs in Fig. 9(c) grows quickly. MLS becomes high, as shown in Fig. 9(c).
 V. C ONCLUSION [11] J. B. Kenney, “Dedicated short-range communications (dsrc) standards
in the united states,” Proceedings of the IEEE, vol. 99, no. 7, pp. 1162–
In this paper, we introduce a novel dynamic mobility- 1182, 2011.
based clustering scheme for VANETs. In our scheme, vehicles [12] Z. Zhang, A. Boukerche, and R. Pazzi, “A novel multi-hop clustering
scheme for vehicular ad-hoc networks,” in Proceedings of the 9th ACM
are clustered in a single-hop based cluster limited by a International Symposium on Mobility Management and Wireless Access,
predetermined value ”Safe Distance Threshold” Dt , where ser. MobiWac ’11. New York, NY, USA: ACM, 2011, pp. 19–26.
Dt ≤ T R. CH is selected as the vehicle which is at the [Online]. Available: http://doi.acm.org/10.1145/2069131.2069135
[13] S. Ucar, S. C. Ergen, and O. Ozkasap, “Multi-hop cluster based ieee
geographical center of a cluster, and CMs are within Dt range 802.11p and lte hybrid architecture for vanet safety message dissemina-
of the CH, moving in the same direction. A new vehicle state, tion,” IEEE Transactions on Vehicular Technology, vol. PP, no. 99, pp.
called temporary cluster head CHt, is proposed in order to 1–1, 2015.
[14] D. Krajzewicz, J. Erdmann, M. Behrisch, and L. Bieker, “Recent
help cluster formation process. CHt vehicle only exists at the development and applications of sumo–simulation of urban mobility,”
beginning of a cluster formation procedure. It changes its state International Journal On Advances in Systems and Measurements,
to CH or CM as soon as the CH is selected. Cluster main- vol. 5, no. 3&4, 2012.
[15] Y. Chen, M. Fang, S. Shi, W. Guo, and X. Zheng, “Distributed
tenance mechanisms are proposed, including cluster merging multi-hop clustering algorithm for vanets based on neighborhood
and leaving a cluster procedure. follow,” EURASIP Journal on Wireless Communications and
Extensive simulations in NS2 with the vehicle mobility in- Networking, vol. 2015, no. 1, pp. 1–12, 2015. [Online]. Available:
http://dx.doi.org/10.1186/s13638-015-0327-0
put from SUMO demonstrate the superior cluster performance [16] V. Vèque, F. Kaisser, C. Johnen, and A. Busson, “CONVOY: a new
of our scheme over LID scheme and VMaSC scheme, in terms Cluster-based routing Protocol for Vehicular Networks,” in Models and
of average cluster number, average CH duration, average CM algorithms for vehicular networks, H. Labiod and A. Beylot, Eds.
ISTE Publishing Knowledge /John Wiley and Sons Inc, May 2013, pp.
duration, CH change rate, number of vehicle state changes, 91–102, chapter 4. [Online]. Available: https://hal.inria.fr/hal-00821742
CM disconnection frequency, and clustering efficiency. The [17] N. Maslekar, J. Mouzna, H. Labiod, M. Devisetty, and M. Pai, “Modified
simulation results show that our proposed clustering scheme c-drive: Clustering based on direction in vehicular environment,” in
Intelligent Vehicles Symposium (IV), 2011 IEEE, June 2011, pp. 845–
provides higher cluster stability even in a high dynamic traffic 850.
scenario. [18] P. Basu, N. Khan, and T. D. Little, “A mobility based metric for
As future work, we aim to investigate the impact of the num- clustering in mobile ad hoc networks,” in Distributed computing systems
workshop, 2001 international conference on. IEEE, 2001, pp. 413–418.
ber of hops on the cluster performance, since our algorithm [19] E. Souza, I. Nikolaidis, and P. Gburzynski, “A new aggregate local
is based on one-hop approach. Moreover, we will implement mobility clustering algorithm for vanets,” in Communications (ICC),
our proposed clustering algorithm for data aggregation and 2010 IEEE International Conference on, May 2010, pp. 1–5.
[20] Z. Wang, L. Liu, M. Zhou, and N. Ansari, “A position-based clustering
data dissemination for recent VANETs applications. technique for ad hoc intervehicle communication,” Systems, Man, and
Cybernetics, Part C: Applications and Reviews, IEEE Transactions on,
R EFERENCES vol. 38, no. 2, pp. 201–208, 2008.
[21] B. Hassanabadi, C. Shea, L. Zhang, and S. Valaee, “Clustering
[1] Y. Z. Hassnaa Moustafa, Vehicular Networks: Techniques, Standards, in vehicular ad hoc networks using affinity propagation,” Ad Hoc
and Applications, 2nd ed. Auerbach Publications, 2009. Networks, vol. 13, Part B, pp. 535 – 548, 2014. [Online]. Available:
[2] S. Al-Sultan, M. M. Al-Doori, A. H. Al-Bayatti, and H. Zedan, “A http://www.sciencedirect.com/science/article/pii/S1570870513002278
comprehensive survey on vehicular ad hoc network,” Journal of network [22] B. J. Frey and D. Dueck, “Clustering by passing messages between
and computer applications, vol. 37, pp. 380–392, 2014. data points,” Science, vol. 315, no. 5814, pp. 972–976, 2007. [Online].
[3] “Etsi tr 102 638 v1.1.1, intelligent transport systems (its); vehicular Available: http://www.sciencemag.org/content/315/5814/972.abstract
communications; basic set of applications; definitions.” [23] M. Togou, A. Hafid, and L. Khoukhi, “Scrp: Stable cds-based routing
[4] Y. Toor, P. Muhlethaler, A. Laouiti, and A. D. L. Fortelle, “Vehicle protocol for urban vehicular ad hoc networks,” Intelligent Transportation
ad hoc networks: applications and related technical issues,” IEEE Systems, IEEE Transactions on, vol. PP, no. 99, pp. 1–10, 2016.
Communications Surveys Tutorials, vol. 10, no. 3, pp. 74–88, Third [24] M. Almalag and M. Weigle, “Using traffic flow for cluster formation in
2008. vehicular ad-hoc networks,” in Local Computer Networks (LCN), 2010
[5] J. Yu and P. Chong, “A survey of clustering schemes for mobile ad hoc IEEE 35th Conference on, Oct 2010, pp. 631–636.
networks,” Communications Surveys Tutorials, IEEE, vol. 7, no. 1, pp. [25] M. Morales, C. S. Hong, and Y. C. Bang, “An adaptable mobility-aware
32–48, First 2005. clustering algorithm in vehicular networks,” in Network Operations and
[6] R. S. Bali, N. Kumar, and J. J. Rodrigues, “Clustering in vehicular Management Symposium (APNOMS), 2011 13th Asia-Pacific, Sept 2011,
ad hoc networks: Taxonomy, challenges and solutions,” Vehicular pp. 1–6.
Communications, vol. 1, no. 3, pp. 134 – 152, 2014. [Online]. Available: [26] K. Abboud and W. Zhuang, “Stochastic modeling of single-hop cluster
http://www.sciencedirect.com/science/article/pii/S2214209614000217 stability in vehicular ad hoc networks,” IEEE Transactions on Vehicular
[7] C. Lin and M. Gerla, “Adaptive clustering for mobile wireless networks,” Technology, vol. 65, no. 1, pp. 226–240, Jan 2016.
Selected Areas in Communications, IEEE Journal on, vol. 15, no. 7, pp. [27] S.-S. Wang and Y.-S. Lin, “Passcar: A passive clustering aided routing
1265–1275, Sep 1997. protocol for vehicular ad hoc networks,” Computer Communications,
[8] G. Yan and S. Olariu, “A probabilistic analysis of link duration in vehic- vol. 36, no. 2, pp. 170 – 179, 2013. [Online]. Available:
ular ad hoc networks,” IEEE Transactions on Intelligent Transportation http://www.sciencedirect.com/science/article/pii/S0140366412002897
Systems, vol. 12, no. 4, pp. 1227–1236, Dec 2011. [28] Network simulator 2 (ns-2), http://isi.edu/nsnam/ns/.
[9] Y. Allouche and M. Segal, “Cluster-based beaconing pro- [29] “Intelligent Transport Systems (ITS);Vehicular Communications; Basic
cess for {VANET},” Vehicular Communications, vol. 2, Set of Applications; Part 2: Specification of Cooperative Awareness
no. 2, pp. 80 – 94, 2015. [Online]. Available: Basic Service,” ETSI EN 302 637-2 V1.3.1, pp. 1–44, Spt 2014.
http://www.sciencedirect.com/science/article/pii/S221420961500011X
[10] A. Daeinabi, A. G. P. Rahbar, and A. Khademzadeh,
“Vwca: An efficient clustering algorithm in vehicular ad hoc
networks,” Journal of Network and Computer Applications,
vol. 34, no. 1, pp. 207 – 222, 2011. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S1084804510001384