Performance Evaluation of Clusterbased TargetTracking Protocols for Wireless Sensor Networks
Aysegul Alaybeyoglu
∗
, Orhan Dagdeviren
†
, Kayhan Erciyes
‡
, Aylin Kantarci
∗∗
Computer Engineering Department, Ege University, Izmir, Turkeyemail:
(aysegul.alaybeyoglu,aylin.kantarci)@ege.edu.tr
†
Computer Engineering Department, Izmir Institute of Technology, Izmir, Turkeyemail:
orhandagdeviren@iyte.edu.tr
‡
International Computer Institute, Ege University, Izmir, Turkeyemail:
kayhan.erciyes@ege.edu.tr
Abstract
—Target tracking is an important application type forwireless sensor networks (WSN). Recently, various approaches [111] are proposed to maintain the accurate tracking of the targetsas well as low energy consumption. Clustering is a fundamentaltechnique to manage the scarce network resources [1219].The message complexity of an application can be signiﬁcantlydecreased when it is redesigned on top of a clustered network.Clustering has provided an efﬁcient infrastructure in manyexisting studies [18]. The clusters can be constructed before thetarget enters the region which is called the static method [14] orclusters are created by using received signal strength (RSS) fromtarget which is called the dynamic method [58]. In this paper weprovide simulations of static and dynamic clustering algorithmsagainst various mobility models and target speeds. The mobilitymodels that we applied are Random Waypoint Model, RandomDirect Model and Gauss Markov Model. We provide metrics tomeasure the tracking performance of both approaches. We showthat the dynamic clustering is favorable in terms of trackingaccuracy whereas the energy consumption of static clustering issigniﬁcantly smaller. We also show that the target moving withGauss Markov Model can be tracked more accurately than theother models.
I. I
NTRODUCTION
Recent technological advancements made sensor nodescheap and readily available for academic and industrial usage.WSNs may consist of thousands of nodes deployed in a largearea. Sensor nodes are suitable for various application typesdue to their sensing and wireless communication capability.Military surveillance, habitat monitoring and target trackingare some of the important types of applications for sensornetworks.In target tracking applications when a mobile target issensed by some of the nodes, its position is calculated bycooperation of these nodes using localization techniques andaggregated data is sent to the sink node. Target trackingapplications can be cluster based [1]–[8] spanning tree based[9], and prediction based [10], [11]. In spanning tree basedalgorithms, nodes which detect the target select a root andconstruct a spanning tree. The tree is conﬁgured while thetarget moves away. Prediction based tracking algorithms aimto estimate the next position of the target based on the currentmoving speed and direction of the target.Clustering is a widely used technique to ease the routingoperation and to manage the scarce resources in WSNs [12]–[19]. In clustered networks, nodes are either classiﬁed ascluster members or cluster heads. Cluster heads are responsiblefor managing the intracluster and cooperating in interclusteroperations. In cluster based target tracking algorithms, membernodes detect the target and send the information to their clusterhead. Cluster heads collect all information from membersand calculate the position of the target by using localizationtechniques. After position of the target is calculated, clusterhead sends the position information to the sink. Reducing theenergy consumption is one of the most important beneﬁts of the cluster based approaches. Cluster based target trackingalgorithms can be further divided into two groups: static [1]–[4] and dynamic [5]–[8] approaches. In static approaches, thecluster and backbone infrastructures are built before the targettracking application starts. On the other hand, the clusters aredynamically constructed while nodes are sensing the target indynamic approaches.In this study, we investigate and evaluate the tracking performance of dynamic and static cluster based target trackingapproaches against various mobility models. To compare thetracking accuracy of two approaches, we measure the missand error ratio. We also evaluate the energy consumptions of dynamic and static approaches and give a general performanceoverview. The rest of this paper is organized as follows: InSection 2, the static and dynamic target tracking algorithmsimplemented are explained. The selected mobility models aredescribed in Section 3. The performance evaluations obtainedfrom simulation results are presented in Section 4. Finally,conclusions and future works are given in Section 5.II. A
LGORITHMS
A. Static Cluster Based Target Tracking Algorithm
In this approach, clusters are formed statically at the time of network deployment so all the member nodes and their relatedleader nodes are deﬁned before the tracking algorithm comesinto play. This cluster ready infrastructure brings simplicityinto target tracking and decreases the energy consumption.Although these are desirable features of this approach, therestrictions on memberships can cause some problems for faulttolerance.
Fig.1 illustrates the general idea in the static cluster basedtarget tracking algorithm. As the target enters the network area,it will be detected by a cluster of nodes in which the targetcurrently presents. The leader node, collects all the sensingreports from its members and by using one of the localizationtechniques it calculates the location, speed and the trajectory of the target. After that, it predicts the future location of the targetand informs the cluster head, closest to the future location of the target, about the oncoming target. When the cluster headreceives this information, it wakes up its members and makesthem ready to detect the moving target. Member nodes sendreceived signal strength values to the new active leader nodeas long as they sense the target. By using the sensing reportsreceived from the member nodes, the new active leader nodecalculates the location, speed and the trajectory of the target.This process continues as long as the target moves and thenodes sense.
predictedfuture locationL1(x,y)L2(x,y)
n1n3n5n4n8n6n2n9n7
Fig. 1. Static Cluster Based Target Tracking Algorithm.
As shown in Fig.1, the target is ﬁrstly detected by theleader node n
1
and by its members n
2
, n
3
and n
4
. n
1
collectsthe sensing reports from its members and calculates thelocations L1(x,y) and L2(x,y) intermittently. By using locationinformation, it predicts the future position after a given periodof time. Having this information, n
1
sends a warning messageto the leader node n
5
that is closest to the predicted futurelocation of the target. This warning message means that thetarget is approaching. After receiving this message n
5
, wakesup its member nodes n
6
, n
7
, n
8
, n
9
and makes them ready todetect the target.Fig.2 shows the ﬂow diagram of the static cluster basedtarget tracking algorithm.
B. Dynamic Cluster Based Target Tracking Algorithm
In this approach, clusters are formed dynamically as theevents occur in the network area. This approach does notimpose any restriction on memberships. For example, a nodecan be a member of different clusters at different times whichmakes this approach more advantageous for minimization of localization errors. Although these are desirable features of this approach, the requirement for leader election mechanismincreases the energy consumption for tracking the target.Fig.3 illustrates the general idea of the dynamic clusterbased target tracking algorithm. As the target enters the
Clusters are formed at the time ofnetwork deploymentActive nodes detect the target andsends sensing reports to leader nodeThe leader node calculates the current,and predicts the next locationof the targetCurrent leader node warns theleader node around thepredicted next location of the target
Fig. 2. Flow Diagram of the Static Cluster Based Target Tracking Algorithm.
network area, it will be detected by some nodes that are closerto the target. By using one of the leader election algorithm[20], a node that is closest to the target is selected as theleader node and the cluster is dynamically formed with theleader node’s one hop neighbors. After forming the initialcluster, the leader node calculates the location, speed and thetrajectory of the target by processing the sensing reports sentby its members. By using this information, it predicts thefuture location of the target and sends a warning message tothe node closest to the target’s predicted future location. Thenode receiving this message, becomes the new leader nodeand forms its cluster with one hop neighbors.
predictedfuture locationL1(x,y)L2(x,y)n1n2n3
n4n5n7n6n11n9n8n10
Fig. 3. Dynamic Cluster Based Target Tracking Algorithm.
As shown in Fig.3, the target is ﬁrstly detected by the nodesn
1
, n
2
, n
3
, n
4
and n
5
. The closest node to the target, n
1
, isselected as the leader node and it forms the initial cluster withits neighbors n
2
, n
3
, n
4
and n
5
which send sensing reportsas long as they sense the target. By processing these reports,the leader node n
1
, predicts the future location and sends awarning message to the node n
6
that is closest to the target’spredicted future location. After receiving this message, the newleader node n
6
, forms its cluster with its one hop neighborsn
7
, n
8
, n
9
, n
10
,n
11
.
Fig.4 shows the ﬂow diagram of the dynamic cluster basedtarget tracking algorithm.
Active nodes detect the targetand selects a leader nodeThe leader node forms its clusterand collects all sensing reportsfrom its membersThe leader node calculates the current,and predicts the next locationof the targetNew cluster is formed dynamicallyaroundthe predicted next location of the target
Fig. 4. Flow Diagram of the Dynamic Cluster Based Target TrackingAlgorithm.
III. M
OBILITY
M
ODELS
Various mobility models were proposed to mimic the natureof mobile nodes in real applications. A hierarchial classiﬁcation of the mobility models is depicted in Fig. 5 [21]. Inthis section, we brieﬂy review the random waypoint, randomdirection and GaussMarkov mobility models. These modelsare chosen because they are commonly used by researchersand their simulations can be handled with ease using
ANSim
simulator [22]. Random waypoint and random direction models are belong to the group of randombased mobility models.In random based mobility models, the nodes choose theirspeed, direction and destination without any restrictions. Onthe other hand, in mobility models with temporal dependency,the velocity of a node at different time slots are correlated.GaussMarkov mobility model is an example of the mobilitymodels with temporal dependency [21]. For each describedmobility model we show the plot of an example trajectorygenerated with the
ANSim
simulator.
A. Random Waypoint Model
Random waypoint model (RWM) [23], [24] is one of the widely used mobility model in WSN simulations. CMUMonarch group implemented RWM for ns2 as the setdesttool. The idea of the RWM is simple. Each node travelsfrom a starting coordinate to a random ending coordinatewith a randomly generated constant velocity. The velocity israndomly picked from [0, V
max
] interval. For greater V
max
values, the nodes move faster since V
avg
= (0 + V
max
)/2,becomes greater. When a node reaches the destination point, itwaits for a T
pause
time before arriving in the next destination.As the T
pause
becomes greater, the total time of the nodes inthe stationary state increases [21]. In this model, nodes movealong a zigzag line. An example of RWM trajectory is shownin Fig. 6.
Fig. 5. Mobility Models.Fig. 6. An Example Trajectory of RWM.
B. Random Direction Model
One of the most important problems of the RWM is thenonuniform distribution of the nodes. The nodes cluster at thecenter of the simulation area as the simulation time elapses. Asthe nodes move to the center, the node density at the border of the simulation area becomes closer to zero causing to a nonuniform node distribution. To overcome this problem, randomdirection model (RDM) is proposed [25]. In RDM, insteadof choosing a random destination, nodes choose a randomdirection which reaches the boundary of the simulation area.When a node reaches the boundary of the simulation area, itwaits for T
pause
time and it chooses a new direction to travel[21]. Fig. 7 shows an example of RDM trajectory.
C. GaussMarkov Model
Random mobility models are simple in nature. They lack capturing realistic behavior of the mobile nodes. In RWM andRDM, unrealistic behaviors such as sharp turns, sudden stops,and sudden accelerations may frequently occur [21]. GaussMarkov Model(GMM) is proposed to prevent these problems[26]. In this model, V
t
+1
is correlated with V
t
where V
t
isthe velocity at time
t
. An example GMM trajectory is shownin Fig. 8.
Fig. 7. An Example Trajectory of RDM.Fig. 8. An Example Trajectory of GMM.
IV. R
ESULTSFig. 9. Error Ratio of STA.
The static cluster based tracking algorithm (STA) anddynamic cluster based tracking algorithm (DTA) are implemented in the
ns2
simulator version 2.31. The clusters andbackbone in STA are created by the algorithm proposed in[18],however other clustering algorithms can also be applied.The backbone of DTA is a simple spanning tree rooted atthe sink. We generated randomly connected networks with 50,
Fig. 10. Error Ratio of DTA.Fig. 11. Comparison of Error Ratios.
100, and 150 uniformly distributed nodes. The mobility modelof target is selected either as RWM, RDM or GMM to measurethe tracking accuracy. The scenario ﬁles of mobility models aregenerated with
ANSim
. Also the speed of the target is varied tomeasure the detection performance of the algorithms. For eachscenario, a lower and an upper bound speed is determined.The speeds are respectively chosen from 5 m/s to 10 m/s, 10 m/s to 15 m/s, 15 m/s to 20 m/s. IEEE 802.11 radioand MAC standards readily available in
ns2
simulator arechosen for lower layer protocols and trilateration is used asthe localization method. In this technique, at least three of thenodes in a cluster must detect the object in order to calculatethe position of the object and the simulation time is 200 s.In our simulations, we used the error and miss ratios as theperformance metrics. If the distance of the calculated positionof a node to the srcin at time
t
is
c
t
and the distance of thereal position to the srcin at time
t
is
r
t
, the error ratio (
e
t
) attime
t
can be calculated as:
e
t
=

c
t

r
t

/
r
t
. To ﬁnd the errorratio, we averaged the measured values during the simulationtime. If the target’s signal at
t
is not collected by at least 3nodes belonging to the same cluster, the target’s location cannot be calculated, thus the target is missed at time
t
. The missratio of the target is the percentage of miss to the total signaldissipated by the target.Firstly, we measured the tracking accuracy of STA and DTAagainst mobility models to compare the effects of RWM, RDM
Fig. 12. Miss Ratio of STA.Fig. 13. Miss Ratio of DTA.
and GMM. The error ratio of the STA and DTA against variousmobility models and speeds are shown in Fig. 9 and Fig. 10.Regardless of the implemented tracking algorithm, the errorratio measured when tracking the target with GMM is smallestand RWM is the highest as seen in Fig. 9, Fig. 10 and Fig.11. Generally, the measured error ratios are very small, at theworst case it is equal to 1.2
%
. On the other hand differenceof miss ratios between mobility models have signiﬁcant valuesand reach approximately 20
%
as shown in Fig. 12, Fig. 13,and 14. The main reason for this difference between mobilitymodels is the sharp turns and sudden accelerations. The missratio difference between models obviously shows that accuratetracking of the target with a random mobility becomes veryhard as the target moves faster.Fig. 11 and Fig. 14 depicts the comparison of the trackingaccuracy of STA and DTA more clearly. The measurementsare collected for speeds 15  20 m/s. Both of error and missratios of STA are higher than those of DTA as seen in Fig.11 and Fig. 14. In random models, the difference betweenerror and miss ratios of algorithms are higher, on the otherhand, algorithms calculate approximate trajectories in GMM.For the target with RWM, approximately 1/4 of the signalsdissipated by the target is missed by STA. In STA, clustersare constructed before the target enters the region whereasthe clusters are dynamically created in DTA by using RSSvalues obtained from the target. For this reason, the chanceof locating the nodes which belongs to the same cluster
Fig. 14. Comparison of Missing Ratios.Fig. 15. Comparison of Trajectories.
which are geographically close to the target is higher inDTA. As mentioned before, in the trilateration technique, atleast 3 nodes of the same cluster must detect the target forlocalization. To illustrate the tracking performance of DTAand STA, a sample scenario with RDM is applied to both of the algorithms. Fig. 15 shows a sample trajectory of the targetand the trajectories produced by STA and DTA for speeds 1520 m/s. The trajectory generated by DTA is very close to thereal trajectory, on the other hand, trajectory of the STA hassome deviations as seen in Fig. 15.Lastly, we compare the energy consumptions of STA andDTA as shown in Fig. 16. In STA, clusters are constructedduring network deployment thus there is a cost of initial energyconsumption of clusters. However DTA has no initial clusterformation but clusters are dynamically constructed when targetis sensed. As the simulation time elapses from 100 s to 500s, the energy consumption of DTA increases linearly with ahigher slope than STA as shown in Fig. 16. The total energyconsumption of DTA reaches to approximately 3 times of theenergy consumption of STA. From the energy consumptionvalues, we can state that the initial energy cost of STA can beseen at 100.s, but dynamic construction of the clusters appliedin DTA is a very expensive operation compared to the energyconsumption of STA can be seen at 500.s.