A Comparative Survey of VANET Clustering Techniques

TL;DR: This paper explores the design choices made in the development of clustering algorithms targeted at VANETs and presents a taxonomy of the techniques applied to solve the problems of cluster head election, cluster affiliation, and cluster management, and identifies new directions and recent trends in the design of these algorithms.

Abstract: A vehicular ad hoc network (VANET) is a mobile ad hoc network in which network nodes are vehicles—most commonly road vehicles. VANETs present a unique range of challenges and opportunities for routing protocols due to the semi-organized nature of vehicular movements subject to the constraints of road geometry and rules, and the obstacles which limit physical connectivity in urban environments. In particular, the problems of routing protocol reliability and scalability across large urban VANETs are currently the subject of intense research. Clustering can be used to improve routing scalability and reliability in VANETs, as it results in the distributed formation of hierarchical network structures by grouping vehicles together based on correlated spatial distribution and relative velocity. In addition to the benefits to routing, these groups can serve as the foundation for accident or congestion detection, information dissemination and entertainment applications. This paper explores the design choices made in the development of clustering algorithms targeted at VANETs. It presents a taxonomy of the techniques applied to solve the problems of cluster head election, cluster affiliation, and cluster management, and identifies new directions and recent trends in the design of these algorithms. Additionally, methodologies for validating clustering performance are reviewed, and a key shortcoming—the lack of realistic vehicular channel modeling—is identified. The importance of a rigorous and standardized performance evaluation regime utilizing realistic vehicular channel models is demonstrated.

Figures

Fig. 3: An illustration of the snaking cluster head chain. The precedence approach may result in chains of subclusters, in which cluster heads are closely grouped together, as seen in the left-hand figure. The boundary nodes are four hops apart in this example. On the right a master cluster head is shown, one of whose members is also a subcluster head. In this scenario, the boundary nodes are only three hops apart. A purely selfish method may not result in the best clustering structure.

TABLE III: Algorithms using Precedence as the Selection Strategy

Fig. 10: Cluster comparisons. Algorithms are sorted clockwise in order of the date of publication. The colour of the ribbon matches the colour of the algorithm that is being proposed, and the ribbon leads to the algorithm to which it was compared. MANET algorithms jut out from the edge of the circle. Diagram created with Circos [112].

Fig. 9: Reaffiliation rate of surveyed algorithms under free space path loss and the experimentally-derived URC channel model. The simplistic model gives an overly optimistic view of performance, and hides disparities between relative performance of clustering methods which are seen when a more realistic model is used.

TABLE I: List of clustering algorithms in this survey

Fig. 6: Cluster lifetime of Hello-based MDMAC and pollingbased RMAC under a simple free-space path loss channel mode, and a more sophisticated urban radio channel model, URAE [91]. This was taken from the simulation survey in [90] (See Figure 5). The disparity of RMAC’s lifetime shows the vulnerability of polling to complex channel phenomena such as fading.

Fig. 1: Lineage of VANET clustering algorithms. LID/HD [18] is noted as the earliest significant VANET clustering scheme, as can be seen from its influence on subsequent algorithms

Fig. 7: Simulators utilised in validation of the surveyed algorithms.

TABLE VII: List of simulators and the publications in which they were first described.

TABLE VI: Gateway selection metrics

Fig. 4: An illustration of the faulty affiliation malfunction. Join messages are lost due to packet collisions or fading; these cannot be detected by an affiliating node, unless the join requires an explicit acknowledgement on the part of the cluster head. In some cases, faulty nodes become isolated from the network, as broadcast-based maintenance mechanisms prevent the faulty node from disconnecting and seeking out other clusters to join. This limits the use of algorithms such as MDMAC [28] in highly dynamic networks.

Fig. 5: Fault affiliation as a percentage of reaffiliation, from the simulation survey in [90]. MDMAC was simulated using three different cluster head selection metrics: Lowest ID, Highest Degree, and LSUF [63]. NL stands for the number of lanes on the highway that comprised the simulation scenario.

TABLE V: Algorithms using Passive Clustering as the Selection Strategy

TABLE IV: Algorithms using Timer as the Selection Strategy

TABLE VIII: Surveyed algorithms and their validation methodologies. “N/S” stands for “Not Simulated”

TABLE II: Algorithms using Weighted Metric as the Selection Strategy

Fig. 8: Channel models utilised in validation of the surveyed algorithms. The two most popular models are Static RTX and unspecified, which makes direct comparison beween algorithms based on performance reported in the literature difficult.

Fig. 2: The basic flow of a clustering algorithm. There are variations on this method, but each algorithm in this survey follows this conceptual flow.

Full text

© 2016 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for
all other uses, in any current or future media, including reprinting/republishing this material for
advertising or promotional purposes, creating new collective works, for resale or redistribution to
servers or lists, or reuse of any copyrighted component of this work in other works.

1
A Comparative Survey of VANET Clustering
Techniques
Craig Cooper
∗
, Daniel Franklin
†
, Montserrat Ros
‡
, Farzad Safaei
∗
, and Mehran Abolhasan
†
∗
ICT Research Institute – University of Wollongong, Wollongong NSW 2500, Australia
email: {cc808|farzad}@uow.edu.au
†
School of Computing and Communications – University of Technology, Sydney, Broadway NSW 2007, Australia
email: {daniel.franklin|mehran.abolhasan}@uts.edu.au
‡
School of Electrical, Computer and Telecommunications Engineering – University of Wollongong, Wollongong
NSW 2500, Australia
email: montse@uow.edu.au
Abstract—A vehicular ad hoc network (VANET) is a mobile
ad hoc network (MANET) in which network nodes are vehicles
– most commonly road vehicles. VANETs present a unique range
of challenges and opportunities for routing protocols due to the
semi-organised nature of vehicular movements subject to the
constraints of road geometry and rules, and the obstacles which
limit physical connectivity in urban environments. In particular,
the problems of routing protocol reliability and scalability across
large urban VANETs are currently the subject of intense re-
search. Clustering can be used to improve routing scalability
and reliability in VANETs, as it results in the distributed
formation of hierarchical network structures by grouping vehicles
together based on correlated spatial distribution and relative
velocity. In addition to the benefits to routing, these groups
can serve as the foundation for accident or congestion detection,
inforomation dissemination and entertainment applications. This
paper explores the design choices made in the development
of clustering algorithms targeted at VANETs. It presents a
taxonomy of the techniques applied to solve the problems of
cluster head election, cluster affiliation and cluster management,
and identifies new directions and recent trends in the design
of these algorithms. Additionally, methodologies for validating
clustering performance are reviewed, and a key shortcoming –
the lack of realistic vehicular channel modelling – is identified.
The importance of a rigorous and standardised performance
evaluation regime utilising realistic vehicular channel models is
demonstrated.
Index Terms—Clustering, VANET, comparative analysis
I. INTRODUCTION
In a Vehicular Ad hoc NETwork (VANET), participating
vehicles are equipped with wireless transceiver which allow
them to exchange data with other neighbouring vehicles, and,
where necessary, to route packets via neighbouring vehicles
to destinations that are not within direct communications
range. One-hop connectivity to external infrastructure is not
necessary, although stationary roadside units may also partic-
ipate in a VANET. Such an architecture potentially enables
the creation of applications ranging from improved traffic
safety and congestion avoidance to in-car information and
entertainment systems.
VANETs operate in a challenging communications environ-
ment, which to date have limited the practical deployment
of the technology. VANETs are particularly susceptible to
the hidden node problem; in addition, they must contend
with limited spectral bandwidth and a highly variable channel
influenced by both stationary and mobile obstructions and
interference sources. In such an environment, infrastructure-
based networks hold a significant advantage over ad hoc
networks: access points allows optimal scheduling of channel
access and distribution of network resources in a relatively
simple manner, at the cost of needing to deploy a large number
of access points throughout the intended coverage area. To
achieve some of the benefits of an infrastructure-based network
without the need for physical infrastructure, researchers have
investigated the idea of clustering in VANETs, whereby a
hierarchical network structure forms in a distributed manner
throughout the network via some sort of clustering algorithm.
Since the introduction of the earliest clustering algorithms
in the late 1980s, a wide range of approaches to the problem
have been proposed in the context of MANETs in general
and VANETs in particular. Each approach is directed toward
different classes of problems and often towards specific appli-
cations envisioned for VANET technology. From the literature,
it is apparent that attempts to validate proposed protocols
often utilise simple channel and mobility models that do
not realistically reflect the VANET environment; they also
frequently compare new VANET clustering techniques to well-
known MANET approaches that have not been designed for
the characteristics of the VANET scenario, resulting in a
favourable comparison but without direct reference to other
state-of-the art approaches. In this paper, we intend to address
these problems in the literature and provide a clear taxonomy
and comparative performance review for all clustering strate-
gies presently employed in VANETs.
A. Structure and Contributions
This review paper makes three significant contributions to
the field of VANET clustering:
1) The main applications of clustering are described, with
general purpose clustering algorithms being considered
separately to application-specific algorithms; this discus-
sion is presented in Section III.
2) The methods by which the major contemporary and
historical algorithms approach the main aspects of the

2
clustering problem are discussed in detail, in particular:
how the cluster head is elected, how unclustered nodes
affiliate with a head, and how cluster heads manage inter-
actions with other clusters. This review suggests a taxon-
omy of VANET clustering techniques which is structured
around these design questions; although VANET clus-
tering taxonomies have previously been described in the
literature, most notably in [1] and [2], these taxonomies
have combined application and design approach within
their classification system. This publication makes a clear
distinction between application and design. The proposed
taxonomy classifies protocols according only to how the
algorithm solves the facets of the clustering problem,
which is a novel approach to the subject. The survey is
presented in Section IV.
3) Finally, the problem of evaluating and comparing the per-
formance of clustering algorithms is considered. Section
V discusses the simulator frameworks, channel models,
and approaches for comparing protocols against compet-
ing algorithms. To the knowledge of the authors, this is
the first such analysis of clustering benchmarking.
A number of potential new directions for clustering research
are identified. The paper concludes with a call for standardised
verification methodologies, and a commentary regarding the
approach to VANET research.
II. CLUSTERING CONCEPTS
A. Development of VANET technology
Serious interest in VANET technology started to develop
in the early 1990s, and has increased in recent years. The
need for standards in this area become apparent after the
emergence of electronic road tolling systems based on a variety
of proprietary active RFID transponder systems. Realising that
this simple concept could be generalised to support a vari-
ety of vehicular communications applications, several major
manufacturers of electronic toll collection systems established
the Dedicated Short-Range Communications (DSRC) Industry
Consortium, which worked towards a common physical layer
standard for short-range vehicular communication. This effort
led to the development of a physical layer based on the ad
hoc mode of IEEE 802.11 operating in the 5.9 GHz band; this
is known as the Wireless Access in Vehicular Environments
(WAVE) / Dedicated Short Range Communications (DSRC)
standard and has been formalised by the IEEE as 802.11p
[3], [4]. Building on this physical layer, a range of standards
for various other parts of the network stack are currently
under development by the IEEE, ISO, ETSI and other bodies
– in particular, the IEEE 1609 working group is developing
standards for security and applications for VANETs built on
top of 802.11p [5]–[8].
Over the last few years, projects such as Keystone Architec-
ture Required for European Networks (KAREN) [9] attempted
to provide a framework whereby policies and plans regarding
vehicular technology could be translated into system specifi-
cations. The EU co-funded Cooperative Vehicle-Infrastructure
Systems (CVIS) project aims to construct a VANET architec-
ture for the provision of a number of services, with considera-
tion of roll-out and public adoption of the technology [10]. The
Cooperative Systems for Intelligent Road Safety (COOPERS)
project aims to develop telematics applications for vehicular
communication systems, and construct a cooperative traffic
management between vehicle and infrastructure [11]. Another
project, SAFESPOT, works toward developing systems for
reoad and vehicle safety [12]. Meanwhile, the NSF Project is
focused on developing safety-specific applications in Vehicular
Networks [13].
VANETs offer exciting opportunities in the areas of traffic
safety and road network efficiency. Cars can avoid collisions
by conversing and exchanging information regarding driver
intention at a level not possible with basic communication
mechanisms such as turn signals. Emergency vehicles can
alert cars ahead, potentially well beyond the human visual
or auditory range, so that a path can be cleared ahead of
the vehicle, thereby reducing response times of ambulences
and police cars. Traffic conditions can be monitored and
congestion alerts issued in real time to enable vehicle flows to
be re-routed around obstructions. When merged with emerging
technologies in driver awareness monitoring, vehicles could
alert a central authority when a driver is tired or intoxicated
or when another vehicle appears to be driving abnormally;
most importantly of all, VANETs offer a form of real-time
accountability when accidents do occur. VANET technology
may also be able to reduce the environmental impact of vehicle
emissions, by selecting routes that result in less fuel consump-
tion and lower emissions [14]. However, a number of technical
and ethical issues arise when considering such applications,
such as security and privacy of a driver, and the ability to
prevent malicious agents from interfering with the network’s
operation, e.g. through modification, jamming or fraudulent
generation of vehicular traffic data. These issues will become
even more critical with the roll-out of autonomous or semi-
autonomous vehicles such as Google Car [15], especially with
respect to accountability when the autonomous navigation
systems fail. Several detailed studies of these issues have
previously been presented in the literature [16], [17].
B. Development of Clustering
A VANET clustering algorithm works by associating mobile
nodes into groups – clusters – according to some rule set,
and selecting a node known as the cluster head (CH) to
mediate between the cluster and the rest of the network in
much the same way as an infrastructure wireless access point.
The specific functions of the cluster head differ depending
on the application, as does the method by which it is selected.
The clustering algorithm used to associate nodes with clusters
should ideally be robust to node mobility and sudden changes
in network and cluster topology, and should provide reliable
end-to-end communication across the VANET.
The earliest notable work on clustering began with the
DARPA packet-radio network [19], the intention being to
autonomously form subnets within a Mobile Ad hoc NETwork
(MANET) to facilitate the distribution of network resources.
This work was built upon by Gerla et al., who proposed the
popular Lowest ID and Highest Degree (LID/HD) clustering
algorithms for MANETs [18]. Mobility Clustering (MOBIC)

3
LID/HD
Sp-Cl
MOBIC PC
C-RACCACBLR WCA
CCP
DMAC UF
LSUF / TC-
MAC
UOFCGDMAC MDMAC
DBC CCA E-Sp-Cl
K-Hop ALM VPC CF-IVC
CMGMCBMAC VWCA
SBCA
CSBP
Fig. 1: Lineage of VANET clustering algorithms. LID/HD [18] is noted as the earliest significant VANET clustering scheme,
as can be seen from its influence on subsequent algorithms
was later presented in [20], attempting to incorporate mo-
bility considerations into the clustering phase. Several other
algorithms including Distributed Group Mobility Adaptive
(DGMA) clustering [21] and MobHiD [22] were later pro-
posed, each designed for clustering in MANETs and demon-
strating progressive improvements in performance under sim-
ple mobility models such as random wayoint. As a result of
this practice of building upon earlier designs, a lineage of
various algorithms developed, which is shown in Figure 1.
As a result of the mobility and channel conditions of
VANETs – which distinguish them from MANET scenarios –
the approaches to clustering have been adapted to these unique
properties. Methods for establishing clusters, detecting and
affiliating with established clusters, and maintaining existing
clusters in VANETs range from channel monitoring, mobility
prediction, machine learning, and security assessment. More
recently, methods of grouping vehicles according to their shape
profile, dimensions, and classification have arisen, e.g. [23],
[24]. Methods of extending cluster lifetime by monitoring
changes in vehicle topology have been investigated e.g. [25].
As VANET research slowly moved away from its MANET
origins, the methodologies whereby a cluster was formed di-
versified. Algorithms were designed with specific applications
in mind, including peer-to-peer (P2P) file sharing and channel
access management. Specific cluster head selection and mem-
ber affiliation schemes were adapted to or invented for these
applications. The diversity in clustering strategies has grown
to the point where adequately classifying these algorithms and
identifying new avenues for research is difficult.
C. Anatomy of a Clustering Algorithm
A series of fundamental procedures are involved in the
formation and maintenance of clusters, which may need to
be repeated depending on the rules of the algorithm and the
mobility dynamics of the network. The general procedural
flow of a clustering algorithm is shown in Figure 2. Nodes
participating in or seeking to participate in a cluster will
typically perform some or all of the procedures described
below, with references to Figure 2.
1) Neighbourhood Discovery: When a vehicle initially
joins the road network and decides that it is willing
to participate in a VANET, its communications system
will be turned on and the node is considered to have
entered a network, initially with only itself as a member.
A node will begin by announcing its existence to its
neighbours through a periodic broadcast, while simul-
taneously gathering similar information from its n-hop
neighbours – either passively by listening for broadcasts,
or actively through neighbour solicitation requests (1).
This information typically includes position information
for the neighbouring nodes, and is stored in a neighbour
table for use by the clustering algorithm.
2) Cluster Head Selection: After gathering data about its
environment, a node will then examine the neighbour
table to find a suitable node to act as its CH (2). The
role of the cluster head varies depending on the clustering
algorithm – it may include routing or relaying functions,
and it may also be responsible for determination of cluster
membership. During this process a node will also assess
its own suitability to be a CH. If the chosen CH is found
within the neighbour table (3), a node will proceed to
step 3; otherwise, if the node itself is best suited to be
CH, it will proceed to step 4.
3) Affiliation: The node will contact the neighbour it deter-
mined to be the optimal CH from its own perspective, and
attempt to become a member of that cluster (4). Some
algorithms may require the addressee to already be an
established CH, while others may allow the addressee
to be an unclustered or regular cluster member. There
may be an additional step where a positive or negative
acknowledgement of the affiliation request is returned to
the joining node, possibly followed by an authentication

4
Scan for
Neighbours
(1)
Found a
suitable
CH? (3)
Assess
neighbour CH
suitability (2)
Affiliate with
CH (4)
Monitor CH
link (5)
Lost CH
link? (6)
NO
YES
Make CH
Announcement (7)
Received Affiliation
Request? (12)
Add new
member to
cluster (13)
Monitor current
members (8)
Lost CM? (9)
Drop member from
cluster record (10)
No more
members?
(11)
YES
YES
NO
NO
NO
Resign from
CH role (14)
YES
NO
YES
Fig. 2: The basic flow of a clustering algorithm. There are variations on this method, but each algorithm in this survey follows
this conceptual flow.
step in the case of algorithms targeted toward security-
sensitive applications. Once a node has become a cluster
member (12,13), it will enter step 5b.
4) Announcement: The node, having determined itself to be
the most suitable CH, may then send out an announce-
ment message to its neighbours to begin the formation
and affiliation process (7). When the node has accrued
cluster members, it proceed to step 5a.
5) Maintenance: This step is different based on whether the
node has become a CH or member:
a) As a Head: the node will poll the members of its
cluster and assess the status of the cluster (8). Several
algorithms have multi-step maintenance processes that
allow clusters to change heads, merge with neighbour-
ing clusters, and track lost links to members, e.g. due
to transient disconnections (9,10). A number of events
may change the state of the cluster: if a CH loses all
of its members (11), the cluster is said to have died
(14), and the node returns to step 1; alternatively, one
cluster may merge with another, and the CH of the
smaller cluster may become an ordinary member of
the new, larger cluster. This is common in algorithms
that place an emphasis on creating large clusters for
increased coverage. In this case, the node will go to
step 5b.
b) As a Member: the node will periodically evaluate its
link to its CH (5), either by waiting for a poll frame
from the CH, or by actively sending “alive” messages.
If the node’s link to its CH fails (6), it will return to
step 1. If the node receives an affiliation request from
an unclustered node, it may withdraw from its parent
cluster to become a CH and continue to step 5a; or in
the case of hierarchical algorithms, it may transition
to a combined state where it is a head of a nested
cluster. In this case, it may perform steps 5b and 5a
simultaneously.
These steps are common across clustering algorithms for
MANETs and VANETs. The key difference between algo-
rithms for the two classes of network is in the methods of
selecting the CH. MANET clustering algorithms use generic
methods of CH selection that typically consider location,
velocity, or node density. In practice, there are many clustering
algorithms that work well in generic unconstrained MANET
scenarios, but which perform poorly in vehicular scenarios due
to the unique channel and mobility dynamics of such environ-
ments. By contrast, VANET-specific clusterng algorithms are
aware of the restrictions on node movement imposed by the
road network and normal traffic flow behaviour, and therefore
utilise information such as traffic mobility metrics and lane
structure which are useful in characterising the position and
behaviour of vehicles on a road. For instance, although two
vehicles may be well within theoretical free-space communi-
cation range, they may be unable to communicate because a
building, large vehicle, or some other obstacle is in the way.
Clustering algorithms that are effective in achieving steps 3
and 5 in a vehicular scenario may therefore be quite different
from those which are used in MANETs.
The question of how to optimally design algorithms to
carry out the affiliation and maintenance steps in a generic
vehicular network is not yet settled. Most published work

Frequently asked questions

Q1. What contributions have the authors mentioned in the paper "A comparative survey of vanet clustering techniques" ?

This paper explores the design choices made in the development of clustering algorithms targeted at VANETs. 

Q2. Why is it necessary to perform low-pass filtering on SNR and RSS metrics?

Due to the time-varying nature of the channel – fast and deep fading due to signal reflections from buildings and other vehi-cles – and transient interference from other localised sources of electromagnetic radiation, RSS or SNR measurements may vary significantly from packet to packet; therefore it is often necessary to perform some sort of low-pass filtering on SNR and RSS metrics to avoid undesirable thrashing of the network topology. 

Q3. What are the criteria for a cluster to have a sufficient trust level?

Routing algorithms employ cluster membership as a requirement for gateway status, while algorithms targeted at securitysensitive applications additionally require gateway nodes to have a sufficient trust level to guard against attack from malicious nodes. 

Q4. Why is PC more likely to be chosen as cluster heads?

Because the architecture waits for data to be sent from the upper layers, vehicles running active data-generating services are more likely to be chosen as cluster heads. 

Q5. What is the role of VANETs in traffic safety?

When merged with emerging technologies in driver awareness monitoring, vehicles could alert a central authority when a driver is tired or intoxicated or when another vehicle appears to be driving abnormally; most importantly of all, VANETs offer a form of real-time accountability when accidents do occur. 

Q6. What is the objective of the criteria used to rank potential cluster heads?

The criteria used to rank potential cluster heads are usually chosen with the objective of mitigating the adverse effects of mobility on communications reliability and throughput, the performance of some particular application, or both. 

Q7. How many wireless vehicular nodes are needed for the evaluation of VANET protocols and applications?

Rigorous experimental evaluation of proposed VANET protocols and applications in urban environments requires the deployment of hundreds or thousands of wireless vehicular nodes. 

Q8. Why is the cluster head index based on its relationship with its neighbourhood?

Since the index representing a node’s suitability9 to lead the entire cluster is based on its relationship with its entire neighbourhood, a node may be identified as the best candidate as a cluster head for a particular node even when its connectivity is quite marginal, because it has good connectivity with the other nodes in the cluster. 

Q9. What are the main factors that affect the accuracy of the results of VANET simulations?

Both the choice of channel model and the choice of simulation framework itself have a significant impact on the accuracy of the results of simulations. 

Q10. What are the main reasons for the need for accurate modelling of channel dynamics?

Numerous authors – for example [87], [104], [105] – have previously demonstrated the need for accurate modelling of channel dynamics for meaningful evaluation of the performance of VANET protocols and applications. 

Q11. What are the main shortcomings of the VANET clustering technique?

While these opportunities are significant, meaningful progress in advancing the state of the art in VANET clustering technique is hampered by a number of significant and fundamental shortcomings in the existing literature, which are summarised below, and which need to be adequately addressed if robust and reliable VANETs are to move beyond simulations and into large scale practical deployment. 

Q12. What is the metric used to determine the degree of mobility of a node?

The sum is then divided by the total number of neighbours to normalise the degree of mobility in the vicinity of the node: therefore, nodes with lower values for this metric are located in a relatively stable environment, indicating that such nodes are good candidates for the cluster head role. 

Q13. How can a large simulation be run in a reasonable period of time?

Despite these limits, very large simulations can be run in a reasonable period of time with a modest amount of commodity computing hardware, with the additional benefit of being able to quickly and easily modify the simulated environment.