MCTA: Target Tracking Algorithm based onMinimal Contour in Wireless Sensor Networks

Jaehoon Jeong, Taehyun Hwang, Tian He, and David DuDepartment of Computer Science & Engineering

University of MinnesotaMinneapolis, MN 55455

Email: {jjeong,thwang,tianhe,du}@cs.umn.edu

Abstract—This paper proposes a minimal contour track-ing algorithm (MCTA) that reduces energy consumption fortracking mobile targets in wireless sensor networks in termsof sensing and communication energy consumption.MCTA

conserves energy by letting only a minimum number of sensornodes participate in communication and perform sensing fortarget tracking. MCTA uses the minimal tracking area basedon the vehicular kinematics. The modeling of target’s kinematicsallows for pruning out part of the tracking area that cannotbe mechanically visited by the mobile target within scheduledtime. So, MCTA sends the tracking area information to onlythe sensor nodes within minimal tracking area and wakes themup. Compared to the legacy scheme which uses circle-basedtracking area, our proposed scheme uses less number of sensorsfor tracking in both communication and sensing without targetmissing. Through simulation, we show thatMCTA outperformsthe circle-based scheme with about 60% energy saving undercertain ideal situations.

Index Terms—Sensor Network, Target Tracking, Energy,Tracking Area, Mobile Target, Vehicle, Kinematics, Circle, Con-tour, Sensing, Communication, Optimization, and Minimization.

I. I NTRODUCTION

The energy efficiency is one of the important research issuesin wireless sensor networks since it determines the lifetimeof the sensor network deployed for the intended applications,such as environmental monitoring, area surveillance, and targettracking. Especially, in the target tracking application,theenergy efficiency is the most important factor as it leads tothe long-lived target tracking. In the target tracking setting,an energy-aware target tracking algorithm not only shouldguarantee the tracking of mobile targets (e.g., enemy tanksor vehicles), but also should maximize the sensor networklifetime using a minimum number of working sensor nodes.The tracking area is defined as the possible region wherethe mobile target can reach from its current position duringsome limited time. The legacy tracking scheme [1], [2] usesthe circle-based tracking area for simplicity. Since the mobiletarget, such as vehicle, moves according to its vehicularkinematics [3], it is impossible for it to reach all the area ofthe tracking circle. We found that we can reduce the numberof working sensor nodes in each tracking area if we use thevehicular kinematics that the mobile target moves accordingto. We try to prune out from the tracking circle the mostunlikely region that the target cannot visit during some limited

time. This makes the tracking area be a minimal-sized areabased on the vehicular kinematics. Only the sensors within theminimal tracking area work for target tracking during somelimited time. Thus, by updating the minimal tracking areacontaining the mobile target during the target’s trajectory, thesensor network based on our scheme consumes less energythan the legacy scheme based on tracking circle. We call ourminimal tracking area theminimal contour.

Our contributions in this paper are as follows:

• The modeling of tracking area based on the vehicularkinematics. Our tracking contour based on the vehicularkinematics is used to select a minimum set of workingsensor nodes for some moment along with the target’strajectory.

• The optimization of tracking contour. We optimize thetracking contour in terms of energy cost by adjusting thelifetime of each contour according to the target’s currentspeed.

• The minimization of communication energy consump-tion. We use both transmission power control and direc-tional antenna to minimize the number of sensors thatreceive the tracking contour information and performsensing.

• The considerations on measurement errors for mobiletarget’s movement. Since the measurements can havesome errors for vehicle’s current position, speed, anddirection, these measurement errors are considered tomake a reasonably larger contour with some confidenceinterval.

The rest of this paper is organized as follows: Section IIdescribes the problem formulation for mobile target tracking.Section III explains the minimal contour tracking algorithm.This section includes the modeling of tracking contour, theoptimization of tracking contour, the minimization of com-munication energy consumption, and the measurement errorhandling. In Section IV, we discuss the implementation issuefor our target tracking algorithm. In Section V, we show thatour contour scheme outperforms the legacy scheme based ontracking circle through simulation. In Section VI, we compareour work with the related works. We summarize our work andshed our future work in Section VII.

II. PROBLEM FORMULATION

We propose an energy-aware target tracking algorithm basedon tracking contour in order to maximize the lifetime of thesensor network. The tracking contour is constructed based onthe vehicular kinematics, which allows a minimal number ofsensors near to the target to work in both communication andsensing.

A. Assumptions and Definitions

We have a few assumptions as follows:• The sensing range is a uniform-disk whose radius isr.• The communication radius is adjustable by controlling

RF transmission power [11], [12].• The RF transmission angle is adjustable by using direc-

tional antenna [13]–[15].• The localization scheme is provided for the sensor nodes

in order to find the position, speed, and direction of thevehicle at any time [4], [5].

We define four terms as follows:Definition 1. Refresh Time. We define the lifetime of of the

tracking area asrefresh time. The old tracking area is replacedwith the new tracking area according to the target’s movementevery refresh time.

Definition 2. Tracking Circle. The tracking circle is thetracking area where the target can visit for its current positionand speed during refresh time. The tracking circle’s radiusisthe multiplication of target’sspeedand refresh time.

Definition 3. Tracking Contour. The tracking contour isthe tracking area where the target can visit for its currentposition, speed and direction during refresh time, consideringthe vehicular kinematics. It prunes out the most unlikely areafrom the tracking circle.

Definition 4. Minimal Contour. The minimal contouris atracking contour for a given target’s speed that allows for theminimization of energy cost spent for target tracking.

B. Main Idea

Our main idea is to minimize the tracking area used todetermine the neighboring sensors that participate in targettracking. The legacy scheme always uses atracking circlesurrounding the mobile target that is modeled as arandomwalk. Though this approach is simple, more than a half of thetracking area based on circle cannot be visited by the targetwithin some limited time [3]. Our scheme uses the vehicularkinematics to prune out the most unlikely area where the targetcannot visit within such small time. Our tracking contour’sshape changes from a circle to a contour (e.g., cone-likeshape) according to the target’s movement state (i.e., stoppingstate and moving state). Our model for tracking contour isrepresented as apolygon approximately including the areawhere the target can reach during refresh time based on thevehicular kinematics. Figure 1 shows two tracking areas: (a)Tracking Circle and (b) Tracking Contour. Letp = (x, y) bethe target’s position vector wherex is x-coordinate andy isy-coordinate. Letm = (v, θ) be the target’s movement vectorwherev is the target’s speed andθ is its direction. We can

−50 0 50−50

−40

−30

−20

−10

0

10

20

30

40

50

X Coordinate [m]

Y C

oord

inat

e [m

]

Vehicle’s Possible Trajectory

Vehicle

Tracking Contour

Tracking Circle

Fig. 1. Tracking Area: Tracking Contour versus Tracking Circle

see that the contour’s area is always the subset of the circle’sarea. So, the contour can allow fewer sensor nodes to trackthe target than the circle; that is, only the sensor nodes withinthe contour whose area is smaller than the circle’s performsensing work, leading to energy saving.

Figure 2 shows the possible trajectories of the vehicle ac-cording torefresh timewhere a new contour is generated fortracking every refresh time. Letone turning timebe the timethat is needed for the vehicle whose speed isv and whoseturning angle is its maximum steering angleφ. Figure 2(a),Figure 2(b), Figure 2(c), and Figure 2(d) show the tracking areafor 1

4turning time, that of2

4turning time, that of3

4turning

time, and that of one turning time, respectively. The outercircle in each figure indicates the tracking area predicted bythe legacy scheme based on circle. Thus, the tracking areais determined withrefresh time, vehicle speed, and turningangle. Thus, since only sensor nodes which belong to thetracking contour smaller than the tracking circle need to turnon their sensing and communication devices, our scheme basedon tracking contour can save more energy than the legacyscheme based on circle.

C. Design Goals

We have three design goals to minimize the energy con-sumption for target tracking: (a) the optimization ofrefreshtime for minimal contour, (b) the minimization ofcommuni-cation costin terms of the number of RF receiving sensors,and (c) each sensor’s localized determination of itswarming-up timeandfinishing timefor sensing.

The refresh timedetermines the size of contour given thetarget’s speed; that is, the bigger the refresh time is, the biggerthe contour is. We need to use the optimal refresh time thatleads to the minimal energy consumption for target tracking.This refresh time is selected as an optimal time, consideringall the energy costs for tracking, such as communication cost,computation cost, and sensing cost.

TheRF transmission power controlanddirectional antennatechnologyare adapted for reducing the communication cost.Because the receiving power consumption is dominant factorin energy cost, we should reduce it. The RF transmission

−40 −30 −20 −10 0 10 20 30 40−40

−30

−20

−10

0

10

20

30

40

X Coordinate [m]

Y C

oord

inat

e [m

]

(a) 1

4Turning Time

−40 −30 −20 −10 0 10 20 30 40−40

−30

−20

−10

0

10

20

30

40

X Coordinate [m]

Y C

oord

inat

e [m

]

(b) 2

4Turning Time

−40 −30 −20 −10 0 10 20 30 40−40

−30

−20

−10

0

10

20

30

40

X Coordinate [m]

Y C

oord

inat

e [m

]

(c) 3

4Turning Time

−40 −30 −20 −10 0 10 20 30 40−40

−30

−20

−10

0

10

20

30

40

X Coordinate [m]

Y C

oord

inat

e [m

]

(d) One Turning Time

Fig. 2. Tracking Contour’s Shape according to Refresh Time

power control and directional antenna technology allow to savereceiving power consumption.

When the sensor nodes turn on and turn off their sensing de-vices can be decided locally in order to save their energy withthe target’s movement information (i.e., position and speed),refresh time and their own position. Refer to Section III-F fordetails.

With the given number of sensor nodes, our objective is tomaximize the sensor network lifetime to satisfy the followingconditions:

• to guarantee the target tracking without missing and• to use the minimal contour appropriate for the target’s

speed in terms of the energy cost in both communicationand sensing.

III. M INIMAL CONTOUR TRACKING ALGORITHM (MCTA)

Assume that the sensor detecting the target can know theposition and speed of the mobile target through the targetlocalization scheme [4], [5]. We define a sensor node dissem-inating the tracking contour information asroot node. When asensor plays a role of root node, it broadcasts the movementinformation of a mobile target.

Algorithm 1 Perform Tracking(contour info)

1: (t, p, v, θ)← Decapsulate Contour Information(contour info){decapsulate the contourinfo into the target’s movement information}

2: ∆T ← Lookup Optimal Refresh time(v){get the optimal refresh time from alook-up table}

3: S ← Compute Minimal Contour Region(p, v, θ, ∆T ){compute the minimal contour’s region with the minimal contour infor-mation sent from the root node with the contour’s center position p, thetarget’s speedv, the target’s direction angleθ, and the optimal refreshtime ∆T .}

4: my position← Get My Position(){my position contains the coordinate of the this sensor node(x, y)}

5: flag ← Am I Inside Minimal Contour(S, my position)6: if flag = TRUE then7: Start Sensing(t)

{this sensor node warms up its sensing devices for sensing}8: Rebroadcast(contour info)

{rebroadcast the new contour’s information to neighbor sensor nodes}9: end if

When the sensor node receives the broadcasted messagecontaining the minimal contour information, it determines

whether it belongs to the minimal contour or not. If thesensor is the member of the new contour, it warms up itssensing devices to prepare for the target tracking and relaysthe message to its neighbor sensor nodes. Otherwise, it justrelays the message to its neighbors.

This section is organized as follows: Section III-A describesthe vehicle’s kinematics and formulates the motion processof the vehicle. Section III-B explains the modeling of track-ing contour based on the vehicular kinematics. Section III-Cdiscusses how to optimize the refresh time for the minimaltracking contour. Section III-E suggests how to expand thetracking contour under measurement errors in the target lo-calization. Section III-F explains how the minimal contourisupdated according to the vehicle’s movement considering theenergy saving related to sensorwarming-uptime.

A. Modeling of Vehicle Motion

We assume that the mobile target is a four-wheeled vehicle.We can define the vehicle motion based on the vehicularkinematics [3], [9]. Figure 3(a) shows the front-end pointPf

and the back-end pointPb of the vehicle that is turning rightwith steering angleφ. Assume that the vehicle’s wheelbase isL that is the distance between the front-end point and back-end point. From the vehicle kinematics [3], we know that thefront-end pointPf is moving on the circle whose radius isRf and the back-end pointPb is moving on the circle whoseradius isRb like in Figure 3(b).Rf and Rb can be obtainedby the following equations [3]:

Rf =L

sin(φ)(1)

Rb =L

tan(φ)(2)

We usePb for vehicle’s position and useRb to make a turningcircle. We can model the vehicle motion by a random vectorM = (X, Y, Θ, V, Φ, V ) in R

6 where(X, Y ) is the positionof the vehicle (Pb), Θ is the orientation (or direction),V is thespeed,Φ is the steering angle, andV is the acceleration [9].We assume that the initial stateM0 = (X0, Y0, Θ0, V0, Φ0, V0)is a Gaussian random vector. LetL be the wheelbase ofthe vehicle. Lets be the time when the vehicle has so farmoved from its first detected time. The driving process thatdetermines the vehicle motion is(Φ, V ). We can compute

L

f R

b R

) , ( b b b y x P

) , ( f f f y x P

c P

=

=

(a) Vehicular Kinematics

f R

b R

c P

b P

f P Front-End

Back-End

(b) Turning Circles of Vehicle

Fig. 3. Vehicular Kinematics for Tracking Contour

the remaining components of M by integrating the followingstochastic differential equations over timet ≥ 0:

dVt = Vtdt (3)

dΘt =Vs

Ltan(Φ)dt (4)

where Vs+∆t = Vs +

∫ ∆t

0

Vt dt and

s← s + ∆t.

(dXt, dYt) = Vs(cos(Θs), sin(Θs))dt (5)

where Vs+∆t = Vs +

∫ ∆t

0

Vt dt,

Θs+∆t = Θs +

∫ ∆t

0

Θt dt = Θs +

∫ ∆t

0

Vs

Ltan(Φ) dt,

and s← s + ∆t.

We can updateXs andYs as follows:

Xs+∆t = Xs +

∫ ∆t

0

Vscos(Θs) dt

Ys+∆t = Ys +

∫ ∆t

0

Vssin(Θs) dt

s← s + ∆t

(6)

The driving processesΦ and V are Gaussian, satisfying thefollowing equations:

dΦt = −αΦt dt + σ dCt

dVt = q dBt

(7)

where(B, C) is a Brownian motion independent ofM0 suchthat (B0, C0) = 0. The constantsα, σ, and q are chosen tosuit particular vehicle motions.

We can consider the vehicle’s backward motion. That is,when the vehicle stops, it can move backward. In this case,we can regard the vehicle’s backward motion to be the sameas the forward motion since it has the same motion equationsabove.

B. Modeling of Tracking Contour

We can make a tracking contour using the vehicular kine-matics discussed in Section III-A. Let(X0, Y0) be the target’scurrent position,Θ0 be the target’s direction, andV0 be target’sspeed. Let∆T be refresh time. Let(X∆T , Y∆T ) be the target’sposition after∆T . We divide target movement into three kinds:(a) Straight movement, (b) Left turning, and (c) Right turning.We can make a polygon representing the tracking contour withthe three styles of movement. Figure 4 shows the procedureconstructing the tracking contour. The straight movement givestwo points in like Figure 4(a). The first point is the target’scurrent position(X0, Y0). The second(X∆T , Y∆T ) is the pointaway from (X0, Y0) by the distance that the target can gowith its current speed and maximum acceleration. The point(X∆T , Y∆T ) can be obtained from Eq. 5. Like in Figure 4(b),the left points can be obtained from Eq. 5 by changing thesteering angle from 0 to maximum steering angle discretelyto the left. In the same way, the right points can be obtainedby changing the steering angle from 0 to maximum steeringangle discretely to the right. The obtained points construct apolygon like Figure 4(d). This polygon is used by each sensorto determine whether it should work for tracking. Only thesensors inside the polygon work, and other sensors continueto be idle. The inside checking is done byRay Crossingsalgorithm [7].

When the refresh time is less than one turning time of thetarget, the tracking contour guarantees the tracking of themoving target without missing. Since the tracking contourcovers all the possible area visited by the mobile target, itguarantees the no-missing tracking. But when the refresh timeis bigger than one turning time, it is very hard to represent atracking contour less than the tracking circle. In fact, in mostcases, as the optimal refresh time is less than one turning timethrough the optimization of refresh time, we need not worryabout the case where the refresh time is bigger than one turningtime. Besides, the fast moving target cannot make a sharp turnto the left or to the right with its maximum steering angle sincethe maximum turning makes the target be overturned.

−40 −30 −20 −10 0 10 20 30 40−20

−10

0

10

20

30

40

X Coordinate [m]

Y C

oord

inat

e [m

]

Target

(a) Case of Straight Movement

−40 −30 −20 −10 0 10 20 30 40−20

−10

0

10

20

30

40

X Coordinate [m]

Y C

oord

inat

e [m

]

(b) Case of Left Turning

−40 −30 −20 −10 0 10 20 30 40−20

−10

0

10

20

30

40

X Coordinate [m]

Y C

oord

inat

e [m

]

(c) Case of Right Turning

−40 −30 −20 −10 0 10 20 30 40−20

−10

0

10

20

30

40

X Coordinate [m]

Y C

oord

inat

e [m

]

(d) Completed Tracking Con-tour

Fig. 4. Construction Procedure of Tracking Contour

C. Optimization of Refresh Time for Minimal Contour

We need to use an optimal refresh time to let the contour-based tracking consume the minimum energy for target track-ing. In this section, we shows how to optimize the refresh timeaccording to the target’s current speed.

Table I shows the notation of parameters used in this paper.The power consumption rates come from the Mica that is oneof Berkeley Motes [10]. Letn be the number of minimal

TABLE INOTATION OF PARAMETERS

Parameter Description∆T Refresh time: unit is [sec]v Mobile target’s speed: unit is [m/sec]

φmax Maximum steering angle: 25◦

φ Steering angle (or turning angle): unit is [◦]d Average trajectory distance which is used to find

out an optimal refresh time∆T : unit is [m]ρ Sensor density that is the average number of

sensor nodes per 1m2 in the surveillance field: unit is [1/m2 ]

n Number of minimal contoursR Maximum communication range: 75[m]

Ptx Energy cost of RF transmitting per second:21[mW]

Prx Energy cost of RF receiving per second: 15[mW]Pcomp Energy cost of computation per second: 16.5[mW]Pwarm Energy cost of warming-up time for preparing

sensing devices in each sensor node per second:15[mW]

Pwork Energy cost of running sensing devices in eachsensor node per second: 10[mW]

Etotal Total energy cost: unit is [mJ]Ttx Time cost of RF transmitting per hop for dissem-

inating the minimal contour information: 0.2[sec]Trx Time cost of RF receiving per hop for receiving

the minimal contour information: 0.1[sec]Tcomp Time cost of computation in a root sensor node

for determining the minimal contour information:0.02[sec]

Twarm Time cost of warming-up time for preparing sens-ing devices in each sensor node: 0.1[sec]

Tsense Time cost of minimum working time for sensingmobile target in each sensor node: 0.5[sec]

Twork Time cost of working time for sensing mobiletarget in each sensor node during the minimalcontour’s lifetime: unit is [sec]

Ttotal Total time cost: unit is [sec]

contours used for tracking given a target trajectory. Letm

be the estimated number of sensor nodes per minimal contourgiven the sensor node density per unit area (i.e.,ρ). The totalenergy costEtotal is the sum of all the operations requiredfor tracking as follows:

Etotal =(PcompTcomp + PwarmTwarm + PworkTsense

+ PworkTwork)mn + PtxTtxn + PrxTrxnρπR2

(8)

To optimize the refresh time∆T minimizing the overallrequired energy for tracking a mobile target, we need toconsider the following. Since the actual trajectory of a mobiletarget is unknown, we cannot see the number of minimalcontours required for tracking the target within the surveillancefield. So, we need to minimize the size of the minimal contourthat is closely related to the energy spent during the generaltarget tracking where the target has directional movementrather than random walk. For the reliable tracking, the refreshtime ∆T should be no less than the sum of computationtime, RF transmitting time, warming-up time, and minimumworking time as follows:

∆T ≥ Ttx + Tcomp + Twarm + Tsense (9)

The shape of the minimal contour is a function of∆T , v, andφmax as follows:

S = f(∆T, v, φmax) (10)

Let d be an average target trajectory distance used to findout an optimal refresh time∆T . We need to optimize∆t

to minimize the total energy spent for the target tracking forthe average trajectory distance; that is, we can formulate ourproblem as follows:

∆T ← arg mint∈R+

Total Energy(t, v, d, φmax) (11)

whereTtx + Tcomp + Twarm + Tsense ≤ t ≤d

v

In order to compute the total energy functionTotal Energy

of (t, v, d, φmax), we need to find the minimal contour’s shapeS corresponding tot by Eq. 10. Next, we find the averagenumber of sensor nodes givenS and the sensor deploymentdistribution, such as a uniform distribution withm sensornodes. Assume that a vehicle with wheelbaseL can turn with

0 1 2 3 4 51

2

3

4

5

6

7

8x 10

5

Refresh Time [sec]

Tot

al E

nerg

y [m

J]

Fig. 5. Searching of Optimal Refresh Time considering TotalEnergyconsumed for Tracking

10001500

20000 10 20 30

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Average Trajectory Distance [m]Target Speed [m/sec]

Opt

imal

Ref

resh

Tim

e [s

ec]

Fig. 6. Optimal Refresh Time according to Vehicle Speed and AverageTrajectory Distance

steering angleφ with speedv. Figure 4(d) shows a minimalcontour’s shape as a polygon wherev = 30 [km/h], L = 2.8[m], andφmax = 25◦.

Algorithm 2 computes the optimal refresh time given thetarget’s speedv, average trajectory distanced, and sensordensity ρ by using the one-dimensional optimization, suchas Golden Section Search algorithm with parabolic interpo-lation [6]. Note that there are many local minimums in the

Algorithm 2 Find Optimal Refresh T ime(v, d, ρ)1: Tmin ← Ttx + Tcomp + Twarm + Tsense

{minimum time needed for minimal contour guaranteeing no missing ofvehicle}

2: Tmax ← d/v{minimum maximum time to cover the average trajectory distance by oneminimal contour}

3: ∆T ← Searching(v, d, ρ, Tmin, Tmax){perform a searching algorithm to find out a refresh time having a globalminimum energy cost within the given range(Tmin, Tmax)}

4: return ∆T{∆T is the optimal refresh time}

refresh time optimization like in Figure 5. The used searchingalgorithm should find out the global optimum among theselocal minimums. Note that the smallest contour made by thesmallest allowable refresh time is not always the minimalcontour in terms of energy cost. In Figure 5, about 1.305[sec]is the optimal refresh time.

Algorithm 3 computes the total energy consumed fortracking given speedv, refresh time t, average trajec-tory distanced, and sensor densityρ by Eq. 8. FunctionsCompute Contour Polygon() and Polygon Area() arecalled to get the polygon approximating to the current contour

Algorithm 3 Compute Total Enery(v, t, d, ρ)

1: n← dd/(vt)e{dxe is the ceiling function ofx}

2: [X, Y ]← Compute Contour Polygon(v, t){this function returns theX and Y vectors representing the polygonapproximating to the contour determined byv and t along with themaximum steering angleφmax}

3: a← Polygon Area(X, Y ){this function returns the area of the polygon represented byX and Yvectors}

4: m← daρe{m is the estimated number of sensors per unit area (i.e.,1m2)}

5: Etotal ← (PcompTcomp + PwarmTwarm + PworkTsense +PworkTwork)mn + PtxTtxn + PrxTrxnρπR2

{Etotal is computed with Eq. 8}6: return Etotal

{Etotal is the total energy needed given refresh timet}

and to compute the area of the contour polygon, respectively[7].

Figure 6 shows the optimal refresh time according to thevehicle’s speedv and average trajectory distanced where theoptimization of refresh time is done. Through Figure 6, wecan see that the optimal refresh time is dominantly affectedby the vehicle’s speed, not the average trajectory distance. InMCTA, the optimal refresh time is chosen for a given vehiclespeed from this graph that is stored as a form oflook-up table.

D. Minimization of Communication Cost

Since the communication cost is dominant factor in energyconsumption in the target tracking, it is worthy to find outhow to reduce such cost in our contour solution. Figure 7shows the communication area for disseminating the contourinformation to the neighboring sensors. Figure 7(a) shows thecommunication area made by full transmission power. It showsthree areas for tracking: (a) the communication circle whoseradius is RF communication range, (b) the tracking contour,and (c) the tracking circle.

For the directional antenna for directional transmission,theroot sensor sends the contour information only towards thesensors belonging to the current contour. The communicationarea can be reduced from the communication circle of Fig-ure 7(a) to the communication cone of Figure 7(b). Throughthis directional transmission, we can modify the energy costfunction of Eq. 8 as follows:

Etotal =(PcompTcomp + PwarmTwarm + PworkTsense

+ PworkTwork)mn + PtxTtxn

+ PrxTrxnρ(πR2)D

2π

(12)

whereD is the angle of directional transmission in radians.D

is selected as a value that can include our contour. The trans-mission area is the cone whose radius is the communicationrange and the internal angle isD. Thus, we can reduce the RFreceiving cost by reducing the number of receiving sensors by1− D

2πtimes, that is, fromρ(πR2) to ρ(πR2) D

2π.

We can minimize the communication cost by letting onlythe sensors within the tracking contour receive the contour

−80 −60 −40 −20 0 20 40 60 80−80

−60

−40

−20

0

20

40

60

80

X Coordinate [m]

Y C

oord

inat

e [m

]

Tracking Contour

Tracking Circle

Communication Circle

(a) Communication Circle by Full TxPower

−100 −50 0 50 100−100

−80

−60

−40

−20

0

20

40

60

80

100

X Coordinate [m]

Y C

oord

inat

e [m

]

Communication Cone

(b) Communication Cone by Tx Power Con-trol

−100 −50 0 50 100−100

−80

−60

−40

−20

0

20

40

60

80

100

X Coordinate [m]

Y C

oord

inat

e [m

]

Minimal Communication Cone

(c) Minimal Communication Cone by bothTx Power Control and Directional Antenna

Fig. 7. Communication Area for Contour Information Dissemination

information. For this purpose, we adopt the transmission (Tx)power control [11], [12] and directional antenna [13]–[15]. Asa result, the communication cone in Figure 7(b) is reduced tothe minimal communication cone like in Figure 7(c). Thus, theenergy cost function can be modified as follows:

Etotal =(PcompTcomp + PwarmTwarm + PworkTsense

+ PworkTwork)mn + PtxTtxn

+ PrxTrxnρ(π(v∆T )2)D

2π

(13)

Finally, the gain of communication energy saving is from Eq.8to Eq.13:

G =PrxTrxnρπ[R2 − (v∆T )2D

2π] (14)

for R > v∆T . Note that in the above formulas, we ignoredanother gain obtained from Tx power reduction for shortercommunication range for simplicity.

E. Handling of Measurement Errors for Target Localization

The target localization is used to locate the most likelyposition of the mobile target with several sensor nodes thatdetected the mobile target at the same time [4], [5]. In orderto estimate the mobile target’s direction and speed, more thantwo localizations are needed where each localization providesa pair of the time and target position. We have assumed so farthat the localization is performed to give these target’s currentposition, speed and direction. However, in reality, since thereare measurement errors in every localization scheme, we needto consider them to make a more realistic tracking contour.

These measurement errors can be modeled as noises thatareGaussianrandom variables [9]. The minimal contour canbe expanded to guarantee containing the tracked target withinthe user-defined confidence interval such as 90%.

Let (Xt, Yt, Θ) be the actual target movement vector whereXt is x-coordinate,Yt y-coordinate, andΘ direction. Let(Xt, Yt, Θt) be a measurement of target movement whereXt

is the measured x-coordinate,Yt the measured y-coordinate,

andΘt the measured direction. We can see that this measure-ment has noise terms as follows:

Xt = Xt + εx

Yt = Yt + εy

Θt = Θt + εθ

(15)

whereεx is a Gaussianrandom variable withN(µx, σ2x), εy

with N(µy, σ2y), and εθ with N(µθ, σ

2θ). Assume that the

user-defined confidence interval isp. Let ξx be the offsetfrom µx satisfying the confidence intervalp. Let ξy be theoffset fromµy satisfying the confidence intervalp. Let ξθ bethe offset fromµθ satisfying the confidence intervalp. Wewill explain the construction procedure of tracking contourunder measurement errors with Figure 8. First of all, like inFigure 8(a), we make a basic tracking contour according toSection III-B. Next, like in Figure 8(b), we make four worst-case contours consideringεx and εy along with ξx and ξy.Next, like in Figure 8(c), we merge these five contours intoa convex hull. Next, considering the direction error like inFigure 8(d), we rotate this convex hull to the left byεθ + ξθ

and also rotate this convex hull to the right byεθ + ξθ. Thesethree convex hulls are merged into a bigger convex hull likein Figure 8(e) usingGraham’s Algorithm[7]. This one is ourfinal tracking contour. It is still smaller than the trackingcircleconsidering the same measurement errors like in Figure 8(f).Note that for the speed measurement error, we just use thegreatest speed corresponding to the given confidence intervalin the same way as the position error and direction error.

F. Update of Minimal Contour

The minimal contour follows the target’s movement chang-ing its refresh time based on the target’s speed. The optimalcontour size is also determined by the average trajectorydistance used for the optimization of refresh time given thetarget’s speed. So, we need to maintain the constant contourshape by changing the refresh time according to the target’scurrent speed with the average trajectory distance of the targetsobserved so far. We can see that the refresh time means the

−60 −40 −20 0 20 40 60

−60

−40

−20

0

20

40

60

(a) Tracking Contourunder No MeasurementError

−60 −40 −20 0 20 40 60

−60

−40

−20

0

20

40

60

(b) Four Tracking Con-tours under PositionMeasurement Error

−60 −40 −20 0 20 40 60

−60

−40

−20

0

20

40

60

(c) Convex Hull repre-senting Merged Track-ing Contour

−60 −40 −20 0 20 40 60

−60

−40

−20

0

20

40

60

(d) Two Convex Hullsunder Direction Mea-surement Error

−60 −40 −20 0 20 40 60

−60

−40

−20

0

20

40

60

(e) Convex Hull repre-senting Merged Track-ing Contour

−60 −40 −20 0 20 40 60

−60

−40

−20

0

20

40

60

(f) Comparison of Fi-nal Tracking Contourand Tracking Circle

Fig. 8. Construction Procedure of Tracking Contour Under Measurement Errors

1 t 2 t 3 t t

2 D warm T

1 D

Fig. 9. Warmimg-up Starting Time of Sensing Devices

lifetime of the current contour when the sensors within itshould continue to work for sensing the target.

Let Tmin be the minimum overhead time to prepare for anew contour in the current contour that is the lower bound forthe refresh time∆T as in Eq. 9; that is,Tmin = Ttx+Tcomp+Twarm + Tsense. Before the target leaves the current contour,the next contour is prepared. That is, if some sensor thatdetects the target the overhead timeTmin before the contour’slifetime ∆T expires, it broadcasts the target’s position andmovement information to its neighbor sensors. The neighborsensors determine whether they will participate in sensingbyperforming Algorithm 1 or not. When the current contour’srefresh time∆T expires, the sensors turn off their sensingdevices except for the sensors that continue to belong to thenext contour.

The starting time of sensing devices is determined con-sidering the movement information message’s timestamp andthe sensor devices’ warming-up time. Table II shows the timevariables needed to compute the warming-up starting time ofsensing devices. We can get the warming-up starting timet2as follows:

t2 = t1 + (D1 − Twarm) (16)

wheret1 is the timestamp of the contour information messagebroadcasted by the current contour’s root node,D1 is the tar-get’s expected travel time from the root node to the computingsensor node, andTwarm is the sensing warm-up time.

IV. D ISCUSSION

A. Computation of Tracking Contour along with OptimalRefresh Time

The complex computation related to optimal refresh time isdone by some powerful sink node outside the sensor network.A table is constructed to have pairs of vehicle speed and cor-responding optimal refresh time. Another table has polygons

TABLE IIT IME VARIABLES FOR COMPUTATION OF WARMING-UP T IME

Parameter Descriptiont1 Time when the contour information was broad-

casted by the root nodet2 Time when the sensor node starts the warming-up

of its sensing devicest3 Time when the sensing devices start the actual

sensingTwarm Time needed for warming-up sensing devices in

sensor nodeD1 Time needed so that the target can reach the sensor

node earliest [17]; that is,D1 = l/v wherel is theEuclidean distance between the target’s startingposition in the current contour and the sensor’sposition andv is the target’s speed

D2 Time difference betweent1 andt2; that is,D2 =t2 − t1

for tracking contour according to the pair of vehicle speed andoptimal refresh time. These two tables are disseminated to thesensors in the sensor network, which are used for our targettracking algorithm in a distributed computing manner. So,since the expensive computation is done in the sink node, thecomputation cost in each sensor is not so high in comparisonwith circle-based tracking algorithm.

B. On-line Classification for Mobile Target

By observing the motion of the tracked target, such as thegreatest turning angle so far, we can figure out a more appro-priate motion process used for constructing a more optimalcontour as in Section III-A. That is, we can use the accuratevehicle motion process according to the estimated vehicle type,such as two-wheeled vehicle, tricycle, four-wheeled vehicle,Reeds-Shepp car, and Dubins car [3].

V. PERFORMANCEEVALUATION

We model the sensor network including sensor and vehicleon the basis of SMPL simulation model along with Matlabwhere SMPL is one of the discrete event driven simulators[8], [16].

A. Simulation Analysis

We define the sensor network lifetime as the time untilat least one sensor node among the sensor nodes on thesurveillance field dies due to the energy exhaustion.

The simulation environment is as follows:

TABLE IIIMEASUREMENT OFNUMBER OF SENSINGSENSORS

Metric Contour (Xt) Circle (Yt) RatioArea 24.1[m2] 137.8[m2] 0.18

Expectation(E) 17 44 0.39V ariance(V ar) 35.6 54.5 0.65

• 10,000 sensor nodes are uniformly deployed in thesurveillance field of 500[m]× 500[m].

• The radius of communication is 75[m].• The vehicle’s speed is 30[km/h] and its maximum turning

angle is25◦.

We simulated to know the number of sensing sensor nodes, thenumber of receiving sensor nodes, and the cumulative energyconsumption according to the vehicle’s movement with thetracking circle and tracking contour, respectively. Figure 10shows these three kinds of performance comparison. LetXt bethe number of working sensor nodes at timet for the contourbased scheme andYt be that at timet for the circle basedscheme. LetE[Xt] be the average number of working sensornodes for the contour based scheme andE[Yt] be that forfor the circle based scheme. LetV ar[Xt] and V ar[Yt] bethe variances ofXt and Yt, respectively. Table III shows thecomparison between two tracking schemes in terms of of thenumber of sensing sensor nodes in our simulation scenario.We can see that the ratio of the expected number of working(i.e., sensing) sensor nodes in contour based scheme to thatin circle based scheme is about 0.25 time, equal to the arearatio (0.18) where the contour’s area is 24.14[m2] and circle’sarea is 137.78[m2]. Therefore, we can conclude that we canreduce the number of working sensor nodes with our minimalcontour scheme, maximizing the sensor network lifetime. Notethat in Figure 10(a), the number of sensing sensors at the firstrefresh time is the same in two schemes. The reason is thatthe tracking contour cannot have enough information for thevehicle’s movement at first, so should use the tracking circle.

The cumulative energy consumptions for two schemes areshown in Table IV. When we do not use both RF transmissionpower control and directional antenna, the performance ratio oftwo tracking schemes is only 0.71; that is, the tracking contourcan improve only 29% of the tracking circle’s performance.The reason is that though 25% of sensors in tracking circle areused in tracking contour, the communication cost that is themajor factor in energy cost is the same in two schemes fromEq. 8. To improve the performance in terms of energy cost, weneed to use the RF transmission power control and directionalantenna discussed in Section III-D. Like in Figure 10(b), wecan reduce a large number of receiving sensor nodes withboth tracking contour and two communication technologies.The tracking contour with two technologies can improve about61% of the tracking circle that also uses the RF transmissionpower control. Note that it is good to use the tracking contouronly with the RF transmission power control and without thedirectional antenna since it still brings the improvement of56%.

0 10 20 30 40 50 60 70 80 9010

15

20

25

30

35

40

45

50

55

60

Time [sec]

Num

ber

of W

orki

ng S

enso

rs

Tracking CircleTracking Contour

(a) Number of Sensing Sensor Nodes Every Refresh Time

0 10 20 30 40 50 60 70 80 900

200

400

600

800

1000

1200

Time [sec]

Num

ber

of R

ecei

ving

Sen

sors

Contour+Power Control+Directional AntennaContour+Power ControlContour+Directional AntennaContour+Full PowerCircle+Power ControlCircle+Full Power

(b) Number of RF Receiving Sensor Nodes Every RefreshTime

0 10 20 30 40 50 60 70 80 900

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

2.2x 10

5

Time [sec]

Cum

ulat

ive

Ene

rgy

Con

sum

ptio

n [m

J]

Contour+Power Control+Directional AntennaContour+Power ControlContour+Directional AntennaContour+Full PowerCircle+Power ControlCircle+Full Power

(c) Cumulative Energy Consumption according to Time

Fig. 10. Performance Comparison between Tracking Circle and TrackingContour

TABLE IVCOMPARISON OFCUMULATIVE ENERGY CONSUMPTION

Method Contour CircleMaximum Transmission Range

(i.e., Full Power) 94166[mJ ] 131770[mJ ]

Directional Antenna 42800[mJ ] N/ARF Tx Power Control 29438[mJ ] 67045[mJ ]

RF Tx Power Control andDirectional Antenna 25988[mJ ] N/A

VI. RELATED WORK

Aljadhai et al. proposed a resource allocation scheme basedon predictive mobility in mobile wireless environments [17].In their paper, the directionality probability was introduced todetermine which cell the mobile target will visit next. Thecell on the direction from the previous cell to the current cellis regarded as the most likely visited cell. Their scheme canbe used for resource allocation in cellular networks havinguser’s mobility profile, but cannot be used for the trackingof a mobile target whose mobility profile is unknown in thewireless sensor network. On the other hand, since our schemeconsiders all the possible tracking area where the mobiletarget can visit mechanically after some time, it guarantees thereliable tracking of the mobile target, such as vehicle, withoutits mobility profile.

Since the papers of [1], [2] model the mobile target asrandom walk, the mobile target can take any direction fromthe current position since the vehicle kinematics are ignored.So, the area where the mobile target cannot visit for sometime belongs to the tracking area. On the other hand, sinceour scheme models the mobile target’s movement based onthe vehicular kinematics [3], only the area where the mobiletarget can visit mechanically belongs to the tracking area.Asa result, we can reduce the number of working sensor nodes ineach tracking area called the minimal contour for the energyefficiency. While the tracking algorithms in [1], [2] focus onthe optimization of the reconfiguration of data collection treefor target tracking, this optimization for tree reconfiguration isout of scope in our paper. Their tracking algorithms can adoptour tracking contour in their tracking algorithm to reduce thenumber of working sensors.

The RF transmission power controlis not only used todetermine the neighboring sensor nodes that can hear thepacket, but also to reduce the communication cost amongclustered sensor nodes [11], [12]. In our tracking algorithm,the transmission power control not only can allow the numberof RF receiving sensor nodes to be minimized, but also canreduce the transmit power for shorter transmission radius.The directional antenna technologyis used for mobile adhoc networks including sensor networks for the parallel com-munication in MAC protocol level [13]–[15]. Our trackingalgorithm uses the directional antenna in order to reduce thenumber of RF receiving sensors.

VII. C ONCLUSION

In this paper, we suggested a target tracking algorithmMCTA using minimal tracking area calledtracking contourthat is based on the vehicular kinematics.MCTA minimizesthe number of working sensor nodes in terms of the com-munication and sensing energy cost during the mobile target’strajectory. We showed that the ratio of tracking contour’s work-ing sensor number to tracking circle’s working sensor numberis proportional to the ratio of the tracking contour’s area totracking circle’s area. This indicates that the reduction of thetracking area leads to the communication and sensing energysaving. We optimize the refresh time for minimal contour

according to the vehicle current speed. Also, in order to reducethe dissemination of tracking contour information within thetracking contour, we used theRF transmission power controland directional antenna, leading to the minimization of thenumber of RF receiving sensors. As our future work, we willimplement our tracking algorithm in real sensor nodes (e.g.,Mica [10]) and test it in our indoor testbed.

ACKNOWLEDGMENT

This work was supported by the Department of ComputerScience and Engineering and Digital Technology Center at theUniversity of Minnesota.

REFERENCES

[1] Wensheng Zhang and Guohong Cao,Dynamic Convoy Tree-Based Col-laboration for Target Tracking in Sensor Networks, IEEE Transactionson Wireless Communications, Vol. 3, No. 5, September 2004.

[2] Wensheng Zhang and Guohong Cao,Optimizing Tree Reconfigurationfor Mobile Target Tracking in Sensor Networks, IEEE Infocom, March2004.

[3] Steven M. LaValle,Planning Algorithms, Cambridge University Press,2006.

[4] Tian He, Chengdu Huang, Brian M. Blum, John A. Stankovic and TarekAbdelzaher,Range-Free Localization Schemes for Large Scale SensorNetworks, ACM Mobicom, September 2003.

[5] Jeongkeun Lee, Kideok Cho, Seungjae Lee, Taekyoung KwonandYanghee Choi,Distributed and energy-efficient target localization andtracking in wireless sensor networks, Elsevier Computer Communica-tions (COMCOM), Vol. 29, No. 13, pp. 2494–2505, August 2006.

[6] D. R. Kincaid and W. W. Cheney,Numerical Analysis: the Mathematicsof Scientific Computing, Van Nostrand, 1991.

[7] Joseph O’Rourke,Computational Geometry in C, 2nd Edition, Cam-bridge University Press, 1998.

[8] M. H. MacDougall, Simulating Computer Systems: Techniques andTools, MIT Press, 1987.

[9] S. J. Maybank, A. D. Worrall and G. D. Sullivan,Filter for Car TrackingBased on Acceleration and Steering Angle, British Machine VisionConference, 1996.

[10] Jason L. Hill and David E. Culler,Mica: a Wireless Platform for DeeplyEmbedded Networks, IEEE Micro, Vol. 22, Nov/Dec 2002.

[11] Vikas Kawadia and P. R. Kumar,Principles and Protocols for PowerControl in Wireless Ad Hoc Networks, IEEE J. Sel. Areas Commun.(JSAC), Vol. 1, pp.76–88, January 2005.

[12] Vikas Kawadia and P. R. Kumar,Power Control and Clustering in AdHoc Networks, IEEE Infocom, March 2003.

[13] Young-Bae Ko, Vinaychandra Shankarkumar and Nitin H. Vaidya,Medium Access Control Protocols Using Directional Antennas in AdHoc Networks, IEEE Infocom, March 1999.

[14] Qingjiang Tian, Seema Bandyopadhyay and Edward J. Coyle, Effect ofDirectional Antennas on Spatiotemporal Sampling in Clustered SensorNetworks, IEEE WCNC, April 2006.

[15] Sensor Network Project for Directional Antenna,https://engineering.purdue.edu/IDEAS/SensorNtwks.html

[16] Matlab of the MathWorks, http://www.mathworks.com[17] AbdulRahman Aljadhai and Taieb F. Znati,Predictive Mobility Support

for QoS Provisioning in Mobile Wireless Environments, IEEE Journalon Selected Areas in Communications (JSAC), Vol. 19, No. 10,October2001.