12
Mobile Netw Appl DOI 10.1007/s11036-011-0311-9 Distributed Active Sensor Scheduling for Target Tracking in Ultrasonic Sensor Networks Fan Zhang · Jiming Chen · Hongbin Li · Youxian Sun · Xuemin (Sherman) Shen © Springer Science+Business Media, LLC 2011 Abstract Active ultrasonic sensors for target tracking application may suffer from inter-sensor-interference if these highly dense deployed sensors are not sched- uled, which can degrade the tracking performance. In this paper, we propose a dynamic distributed sensor scheduling (DSS) scheme, where the tasking sensor is elected spontaneously from the sensors with pending sensing tasks via random competition based on Carrier Sense Multiple Access (CSMA). The channel will be released immediately when sensing task is completed. Both simulation results and testbed experiment demon- strate the effectiveness of DSS scheme for large scale Part of this paper was presented at ChinaCom’10, Beijing, China, Aug. 2010 [1]. F. Zhang · J. Chen (B ) · H. Li · Y. Sun State Key Laboratory of Industrial Control Technology, Zhejiang University, Hangzhou, 310027, China e-mail: [email protected], [email protected] F. Zhang e-mail: [email protected] H. Li e-mail: [email protected] Y. Sun e-mail: [email protected] X. Shen Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, ON N2L 3G1, Canada e-mail: [email protected] sensor networks in terms of system scalability and high tracking performance. Keywords wireless sensor networks · target tracking · sensor scheduling 1 Introduction Wireless sensor networks (WSNs) have been consid- ered as a promising technique for area surveillance applications [2, 3], and target localization/tracking is essential for these applications [4, 5]. Many approaches have been proposed for target tracking within WSNs [68, 15]. According to target behavior, the previous works can be categorized into two classes: coopera- tive [68] and non-cooperative [9, 15]. A cooperative target is part of the network and emits certain forms of physical signals that reveal its presence or report its own identification. In the non-cooperative scenario, however, there exists no information exchange between the target and the network infrastructure. Therefore, sensor nodes need to detect and identify the target actively by emitting energy and measuring the feed- back. Our previous work [9], a tracking system aim- ing at non-cooperative targets, utilized passive infrared sensors for target detection and the active ultrasonic sensors for ranging. In non-cooperative tracking systems using active ultrasonic sensors, Inter-Sensor Interference (ISI) problem is frequently presented if ultrasonic signals of different sensors are emitted within a short time inter- val. When nearby active sensors work simultaneously at same frequency, the ultrasonic receiver may receive the direct or reflected chirp from other transmitter

Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

  • Upload
    others

  • View
    5

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw ApplDOI 10.1007/s11036-011-0311-9

Distributed Active Sensor Scheduling for Target Trackingin Ultrasonic Sensor Networks

Fan Zhang · Jiming Chen · Hongbin Li ·Youxian Sun · Xuemin (Sherman) Shen

© Springer Science+Business Media, LLC 2011

Abstract Active ultrasonic sensors for target trackingapplication may suffer from inter-sensor-interferenceif these highly dense deployed sensors are not sched-uled, which can degrade the tracking performance. Inthis paper, we propose a dynamic distributed sensorscheduling (DSS) scheme, where the tasking sensor iselected spontaneously from the sensors with pendingsensing tasks via random competition based on CarrierSense Multiple Access (CSMA). The channel will bereleased immediately when sensing task is completed.Both simulation results and testbed experiment demon-strate the effectiveness of DSS scheme for large scale

Part of this paper was presented at ChinaCom’10, Beijing,China, Aug. 2010 [1].

F. Zhang · J. Chen (B) · H. Li · Y. SunState Key Laboratory of Industrial Control Technology,Zhejiang University, Hangzhou, 310027, Chinae-mail: [email protected], [email protected]

F. Zhange-mail: [email protected]

H. Lie-mail: [email protected]

Y. Sune-mail: [email protected]

X. ShenDepartment of Electrical and Computer Engineering,University of Waterloo, Waterloo,ON N2L 3G1, Canadae-mail: [email protected]

sensor networks in terms of system scalability and hightracking performance.

Keywords wireless sensor networks · target tracking ·sensor scheduling

1 Introduction

Wireless sensor networks (WSNs) have been consid-ered as a promising technique for area surveillanceapplications [2, 3], and target localization/tracking isessential for these applications [4, 5]. Many approacheshave been proposed for target tracking within WSNs[6–8, 15]. According to target behavior, the previousworks can be categorized into two classes: coopera-tive [6–8] and non-cooperative [9, 15]. A cooperativetarget is part of the network and emits certain formsof physical signals that reveal its presence or reportits own identification. In the non-cooperative scenario,however, there exists no information exchange betweenthe target and the network infrastructure. Therefore,sensor nodes need to detect and identify the targetactively by emitting energy and measuring the feed-back. Our previous work [9], a tracking system aim-ing at non-cooperative targets, utilized passive infraredsensors for target detection and the active ultrasonicsensors for ranging.

In non-cooperative tracking systems using activeultrasonic sensors, Inter-Sensor Interference (ISI)problem is frequently presented if ultrasonic signals ofdifferent sensors are emitted within a short time inter-val. When nearby active sensors work simultaneouslyat same frequency, the ultrasonic receiver may receivethe direct or reflected chirp from other transmitter

Page 2: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

Fig. 1 Inter-sensor interference in ultrasonic ranging

and results in erroneous sensor readings, consequentlyleads to unacceptable estimation results. Figure 1 showsan example of ISI, where Sensor 3’s ranging processperiodically collides with that of Sensor 1. Sensor 3 isalways interfered by sensor 1 and can not get accu-rate measurement signal. An occasional ranging taskinitiated by sensor 2 is interfered by the signal fromSensor 3. Erroneous ranging measurements are evenmore harmful than missing the reflected signal, sincethe target estimator accepts the measurement with highconfidence. Therefore, sensor scheduling is needed toensure that, during any time, only one sensor in an ISIregion can work to detect the target.

Sensor scheduling, also referred as sensor selectionor sensor management, concerns with turning on theright sensors at the right time to achieve desirableperformance with minimal energy consumption. Someprevious works have addressed the sensor schedul-ing problem for different tracking systems in WSNs[10–14]. However, most of them mainly study thetradeoff between tracking performance and energyconsumption. In [10], this problem is formulated as apartially observable Markov decision process, andMonte Carlo method is developed using a combina-tion of particle filters for belief-state estimation andsampling based Q-value approximation for lookahead.Adaptive sensor activation [11] selects the next taskingsensor and its associated sampling interval based onthe prediction of tracking accuracy and energy cost.Priority list sensor scheduling [12] facilitates efficientdistributed estimation in sensor networks, even in thepresence of unreliable communication, by prioritizingthe sensor nodes according to local sensor schedulesbased on the predicted estimation error. In [13], ascheduling scheme is proposed to improve scalabilityand energy-efficiency of localization by limiting thenumber of active nodes participating in localization,while maintaining sufficient coverage to ensure accept-able location error. However, all the methods men-tioned above can not be applied directly to the ISIproblem for active sensors.

In [15], two sensor scheduling schemes for activesensors are proposed. In the periodic sensor scheduling(PSS) scheme, each ultrasonic sensor detects the target

in turn, within the predefined time slots assigned to it. Acritical drawback of this sensor scheduling scheme is theexistence of empty measurement, where a scheduledsensor may not be in the vicinity of the target. Asa result, the system expects lower tracking accuracyand wastes of energy. Whereas in the adaptive sensorscheduling scheme, the next tasking ultrasonic sensoris selected adaptively according to the state predic-tion of the target. In this scheme, each node needsto know the positions of its neighbors, and the sensorselection process is very computational intensive, sincethe current tasking node takes the complex calculationfor node selection. The considerable computation timemay cause delay to the next step sensing, deterioratetracking accuracy and even lead to target loss. In addi-tion, due to the distributed nature of WSNs, the previ-ously mentioned scheduling schemes are less applicablesince scheduling has to be performed in a distributedway to ensure scalability.

In this paper, we propose a distributed sensorscheduling (DSS) scheme for target tracking in ul-trasonic sensor network. The tasking sensor node iselected spontaneously from the sensor candidates ina distributed way via random competition based onCarrier Sense Multiple Access (CSMA). As soon as thesensing task is completed, the channel is released im-mediately for other sensor nodes with pending sensingtasks. It is demonstrated the DSS scheme is effectivefor large scale sensor networks and robust to dynamictopology changes.

The remainder of the paper is organized as follows.The system model and problem setup are introduced inSection 2. Section 3 presents the DSS scheme. Sections4 and 5 evaluate the scheme with extensive simulationand testbed experiment, respectively. Finally, conclu-sions and future work are given in Section 6.

2 Modeling and problem formulation

Typically, each sensor node in WSNs has short sens-ing range and the on-board processor is limited bothin memory and processing speed. In our trackingsystem, each sensor node is built with one PassiveInfraRed (PIR) sensor, multiple ultrasonic rangingsensors and one processing/communication board. Thesensor nodes only take charge of detecting/sensing thetarget and transmitting the sensing results to the centralcomputer. Noise corrupted measurements may causesolving the tracking problem with insufficient accu-racy. Therefore, the Kalman Filter (KF) is used toestimate the state of our dynamic system with noisymeasurements. The current state X(k) of the system is

Page 3: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

calculated recursively from the previous state X(k − 1)

and a new measurement zi(k). An error covariancematrix P is calculated along with X for demonstratingthe accuracy of the estimation.

In this paper, we focus on single target tracking of a2D space and the system is set as X = [x, y]T . The state-space model of the system can be written as

X(k) = FX(k − 1) + w(k)

where X(k) is the state of the target at k-th measure-ment time, F models the system dynamics, and it isassigned as identity matrix due to the unknown targetmotion pattern, w(k) is the variable representing un-modeled process noise or higher order motion.

If sensor i is used for the k-th measurement zi(k) ofthe target, the measurement model is given by

zi(k) = hi(X(k)) + vi(k)

where vi(k) is measurement noise in sensor i. Both wand vi are assumed to have white noise properties withzero mean and their covariances are Q and R j respec-tively. hi is the measurement function and following theultrasonic ranging model, it can be calculated as

hi(X(k)) = ||X(k) − Xi||

where Xi denotes the coordinates of sensor i. Thus, themeasurement process is non-linear with respect to X inour case, and traditional KF is no longer applicable. AnExtended Kalman Filter (EKF) can be applied with thefollowing linearization:

Hi(k) = ∂hi(X(k))

∂X

=[

x− xi√(x − xi)2 + (y− yi)2

,y − yi√

(x− xi)2 + (y− yi)2

]

X(k) and P(k) constitute the state of the EKF andthe recursion is comprised of two phases: (a) a pre-diction phase during which the current state is esti-mated based on the previous state according to thestate-space model, and (b) a correction phase whichis used to refine the predicted state based on the newmeasurements.

KF prediction phase Given the previous state [X(k −1), P(k − 1)], the predicted state [X̄(k), P̄(k)] is

X̄(k) = FX(k − 1)

P̄(k) = FP(k − 1)FT + Q

EKF correction phase With the new measurementzi(k) and the predicted state [X̄(k), P̄(k)], the new state[X(k), P(k)] is refined by

K(k) = P̄(k)Hi(k)T [Hi(k)P̄(k)Hi(k)T + Ri]−1

X(k) = X̄(k) + K(k)[zi(k) − hi(X̄(k))]

P(k) = [I − K(k)Hi(k)]P̄(k)

where K(k) is Kalman Filter gain.For the sensor scheduling problem, we only concern

how to sense the target efficiently via the sensor nodesnegotiation. We first define some notations in Table 1.The definition of function f is:

f (t) =⎧⎨⎩

i, select node i as the tasking node at time t

0, otherwise(1)

In order to avoid ISI between ultrasonic sensors,only one sensor node is tasked to actuate its ultrasonicsensors for range measurement each time. And thetime difference between two successive measurementepoches should be larger than Td. Simulation resultsin [9] reveal that the tracking performance can benefitfrom higher sampling frequency. With view to the exis-tence of empty detection when the scheduled sensor isnot in vicinity of the target, we define another functiong(t) to represent the effective scheduling:

g(t) =⎧⎨⎩

1, f (t) �= 0 and ||x f (t) − x(t)|| < R

0, otherwise(2)

Table 1 Notation definitions

Symbol Definition

| · | | · | represents the cardinality of a set.V Set of the sensor nodes, i.e., V={1,2,...,|V|}.[0, T] Duration that the target in the monitored area.x(t) Position of the target at time t, and t ∈ [0, T].xi Position of sensor node i, and i ∈ V.Td Die-out time of the ultrasonic wave in a ranging

operation.R Sensing range of the ultrasonic sensors.f (t) Function that record the process of the scheduling

during [0, T], the detailed definition is given inEq. 1 and f : [0, T] → V ∪ {0}.

N Total number of nonzero element in f (t), meansthe number of the selection.

ni The ith nonzero element in f (t), means the ithtasking node. So i = 1, 2, ..., N, ni ∈ V.

ti Time of the ith tasking node selection, i.e, f (ti) = ni.

Page 4: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

Therefore, the ISI problem can be represented by anoptimization form:

maximum1T

∫ T

0g(t)dt

subject to ti − ti−1 ≥ Td

i = 2, 3, ..., N (3)

3 Distributed sensor scheduling scheme

By considering the ranging operation of a sensor nodeas the occupation of a shared channel, the ISI problemamong active ultrasonic sensors in WSNs can be con-verted to the problem of multiple access in a sharedchannel. Hence, the MAC protocols can be used tosolve the ISI problem.

A common MAC paradigm is TDMA (time-divisionmultiple access) which schedules all nodes to transmitmessages at different times. The PSS scheme men-tioned above is kind of this fashion. However, TDMAoften requires a central unit to find a collision-freeschedule and developing an efficient schedule with highchannel utilization is very difficult. Meanwhile, it usestopology information as a basis for access schedulingand needs clock synchronization among neighbors, thuslacks efficiency and scalability when the networks sub-ject to frequent topology changes.

Another classic MAC protocol is CSMA, which isa probabilistic protocol. In a nutshell, CSMA verifiesthe absence of other traffic before transmitting in ashared channel. It is popular because of its simplicity,scalability and robustness. It does not require clocksynchronization and global topology information, thusis able to handle dynamic topology changes which gen-erally occur in WSNs without extra operations. Becauseof these features, sensor nodes negotiate with eachother using CSMA is our proposed scheme.

The main idea of DSS scheme is that when a nodehas a pending ultrasonic sensing request, it checks ifthere is already an occupation announced by othernode in recent σ millisecond. If no occupation isrecorded, the node broadcasts its own message toannounce occupation in the upcoming σ millisecondand then conducts target range sensing. Otherwise, thenode waits a bounded random time for next roundoccupation. Obviously, the minimal σ should be equalto Td . The detailed procedure is summarized below:

– a) In initial state, all the nodes are in sleep mode;when a mobile target enters the monitored area,sensor nodes close to the target will be activated

by the PIR sensor. As the target moves along thetrace in the area, some activated sensor nodes maygo back to sleep mode because the target moves outof its sensing region. Also, there are some newlyactivated sensor nodes;

– b) Once a sensor node is activated, it will calculatea random Tbackoff, and then start its delay timer withinterval Tbackoff. Let

Tbackoff ∈ [Tmin, Tmax] and Tmin = Td

– c) As soon as the node with the smallest Tbackoff

triggers its delay timer, it will broadcast a DETECTmessage to the other activated nodes and becomethe tasking node to sense the target. Any node thatoverhears the DETECT message will immediatelygive up its declaration as the tasking sensor nodeby restarting the delay timer with a new randomTbackoff. The timerdelay in the tasking sensor node isalso restarted with new random Tbackoff.

Without loss of generality, it is assumed there is notransmission delay between sensor nodes. Therefore,the DETECT message broadcasted by the selectednode can be used as a locally synchronization signalto make all the activated nodes start their delay timerat the same time. It can be seen that the computationburden is distributed among the activated sensor nodesand the sensor selection is totally distributed. Anotheradvantage of the DSS scheme is that each node does notneed to know the positions of its neighboring nodes andthus conserving the limited memory for other process-ing. Therefore, DSS is well suited for implementationin WSNs.

In theory, it is almost impossible that more thanone sensor nodes broadcast DETECT message at thesame time because the Tbackoff is randomly selected. Inpractice, however, the Tbackoff can only be selected froma finite discrete set due to quantization constraint. Forone sensor node, we have

Tbackoff ∈ {T0, T1, . . . , TM−1, TM}

T0 = Tmin, TM = Tmax

Ti − Ti−1 = �t, i = 1, 2, . . . , M

where �t is a constant value denoting the resolution ofthe delay timer which depends on the sensor nodes andit equals to 1ms in our system. In order to avoid the sit-uation that more than one nodes sense the target at thesame time, the node with the smallest ID will be chosenas the tasking node among the nodes whose delay timerare triggered simultaneously, and Algorithm 1 is thepseudo-code for the DSS scheme. The parameter M

Page 5: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

Algorithm 1 Distributed sensor scheduling algorithm1: Input: the trace of the target x(t)2: Output: the sensor scheduling results f (t)3: Initialization: f (t) = 0, t ∈ [0, T];4: //det_send means broadcast the DETECT message.5: //det_recv( j) means receive the DETECT message from

node j.6: for target moves in the monitored area do7: if target moves into the sensing range of node i then8: sensor i is activated from the sleep mode;9: end if

10: for each activated node i do11: generate a random Tbackoff;12: start the delay timer with Tbackoff;13: wait for det_send or det_recv( j);14: if not det_send and det_recv( j) at time t then15: stop the delay timer;16: generate a new random Tbackoff;17: start the delay timer with Tbackoff;18: set f (t) = j;19: end if20: if det_send and not det_recv( j) at time t then21: sensing the target;22: generate a new random Tbackoff;23: start the delay timer with Tbackoff;24: set f (t) = i;25: end if26: if det_send and det_recv( j) at time t then27: if i > j then28: stop the delay timer;29: generate a new random Tbackoff;30: start the delay timer with Tbackoff;31: set f (t) = j;32: else33: sensing the target;34: generate a new random Tbackoff;35: start the delay timer with Tbackoff;36: set f (t) = i;37: end if38: end if39: end for40: if target moves out of the sensing range of node i then41: sensor i goes back to sleep mode;42: end if43: end for

denotes the size of candidate set for Tbackoff. Larger Mindicates that sensor nodes are less possible to collidewith each other, but the expected backoff time may belarger and result in sensing latency.

As described above, the DSS scheme employs a quitesimple strategy to avoid collision where node ID is usedto break ties when multiple nodes trigger their delaytimer simultaneously. However, the strategy is not soefficient, especially when M is small. This is becausethe node with smaller ID is preferred to be selected

with this strategy and small M always intensifies thisinclination. Consider node i among L activated nodes,which i is the k-th smallest number in the L node IDs,1 ≤ k ≤ L. Since node i is selected to be the taskingnode if and only if when its Tbackoff is the smallest andthe IDs of other nodes which have the same Tbackoff

are all larger than i, the probability Pr(k) that node iis selected can be expressed as follows:

Pr(k) =M∑

i=0

1M + 1

(M − iM + 1

)k−1( M + 1 − iM + 1

)L−k

= 1(M + 1)L

M∑i=0

(M − i)k−1(M − i + 1)L−k

= 1(M+1)L

M∑i=0

[(M− i+1)L−1

(1− 1

M− i+1

)k−1

]

(4)

We can then obtain the gap between the smallest IDnode and the largest one as follows:

Pr(k = 1) − Pr(k = L)

= 1(M + 1)L

M∑i=0

[(M − i + 1)L−1 − (M − i)L−1]

= 1M + 1

(5)

Clearly, according to Eq. 4, Pr(k) decreasesmonotonically with increasing k, i.e., node with smallerID has higher probability to become the tasking node.Meanwhile, Eq. 5 indicates that decreasing M willenlarge the inclination. Figure 2 shows a visualizedsimulation example.

High selected probability of specific node leads torepeated selection of that node which always resultsin large accumulate error. Therefore, we can improvethe DSS scheme by modifying the collision avoidancestrategy to diminish the accumulate error. The keyidea is to use random value to break ties. When anode triggers its delay timer, it will generate a randomvalue and broadcast this value to other activated nodeswith the DETECT message. If it happens that multiplenodes trigger their delay timer simultaneously, the nodewith the smallest random value will be chosen as thetasking node. With this strategy, each node has thesame chance to be the tasking node, which furtherreduces the accumulate error to a bare minimum.

Page 6: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 200

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

Node ID

Sel

ecte

d P

roba

bilit

yM=5, L=20

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 200

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

Node ID

Sel

ecte

d P

roba

bilit

y

M=15, L=20

Fig. 2 Selected probability for all nodes with different M in DSSscheme

4 Simulation results

We evaluate the proposed DSS scheme with both sim-ulation and testbed implementation. We compare thetracking performance of the DSS scheme with the PSSscheme and also validate the efficiency of the improvedDSS scheme.

In the simulation, the monitored area is modeled asa square field. The Random WayPoint mobility model(RWP) [16] is used to generate movement traces of themobile target. An estimation error at one point in thetrace is defined as the Euclidean distance betweenthe estimated position and corresponding true position

e(t) = ‖x(t) − x̃(t)‖ (6)

where x(t) is the ground truth position vector and x̃(t) isthe estimated position.

The mean tracking error is defined as the averagederror of all the estimated points in the trace

e = 1(tK − t0)

�Ki=1(ti − ti−1)e(ti) (7)

Table 2 Default simulation setup

Parameter Description

Field area 6 × 6 m2

Sensing noise model N(0, 0.0025)

Number of sensor nodes 12, uniformly deployedTarget velocity U(0.3, 0.7) m/sPIR range and angle 0 ∼ 3 m, ±π

Ultrasonic range and angle 0 ∼ 3 m, ±π

Td 30 msM 15

where t0 and tK are the starting time and ending time,respectively, and K denotes the total numbers of esti-mations.

The presented results are averaged over 50 inde-pendent simulation runs for high confidence. Table 2illustrates the default simulation setup. For the sake ofclarity, we assume that the sensing angle of PIR sensorand ultrasonic sensor are both ±π , i.e., that the sensingmodel is a circle.

We compare the PSS scheme with the proposedDSS scheme. Figure 3 shows the target moving tracesestimated by PSS and DSS. It can be seen that theestimated trace generated by DSS is closer to the trueone, and the corresponding tracking error of these twoestimated traces are 0.1131 and 0.0575 m, respectively.For a larger network, the tracking results are shown inFig. 4. The monitored area is 12 × 12 m2 with 48 sensornodes deployed. It can be seen that the DSS schemestill performs well under such conditions. On the otherhand, PSS scheme almost lost the target during the firstcorner and at the end of the trace because of the empty

0 1 2 3 4 5 60

1

2

3

4

5

6

x−coordinate(meter)

y−co

ordi

nate

(met

er)

Tracking Results

Sensor NodesTarget Enter PointTarget Exit PointGround Truth of Moving TraceEstimated Moving Trace of PSSEstimated Moving Trace of DSS

Fig. 3 Simulated trajectories of PSS and DSS schemes

Page 7: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

0 2 4 6 8 10 120

2

4

6

8

10

12

x−coordinate(meter)

y−co

ordi

nate

(met

er)

Tracking Results for Large Network

Sensor NodesTarget Enter PointTarget Exit PointGround Truth of Moving TraceEstimated Moving Trace of PSSEstimated Moving Trace of DSS

Fig. 4 Simulated trajectories for larger network

detections. The corresponding tracking error of PSSand DSS are 0.3651 and 0.0534 m, respectively.

4.1 Impact of the number of sensor nodes

We compare the DSS scheme with PSS scheme undera different setting of sensor quantity, ranging from 6to 30. Figure 5 shows that (i) the DSS scheme out-performs the PSS scheme; and (ii) with the increasenumber of deployed sensor nodes, the tracking error

6 9 12 15 18 21 24 27 300

0.20.40.60.8

11.21.41.6

Number of Sensor Nodes

Max

Err

or(m

eter

)

Impact of the Number of Sensor Nodes

Maximum Error of PSSMaximum Error of DSS

(a) Impact of node number on max error

6 9 12 15 18 21 24 27 300

0.05

0.1

0.15

0.2

0.25

0.3

Number of Sensor Nodes

Mea

n E

rror

(met

er)

Mean Error of PSSMean Error of DSS

(b) Impact of node number on mean error

Fig. 5 Impact of node number on tracking error

for both schemes are reduced, and the error convergesto a constant when the node number is very large.This is because the uncovered area is large when lessnumber of nodes are deployed, which leads to highprobability of losing the target. On the other hand, thearea coverage is saturated when the node number islarge enough, so that any more sensor deployment doesnot help in terms of tracking performance.

4.2 Impact of the scale of ROI

ROI means the region of interest, i.e., the monitoredarea. We compare the performance of both schemesunder different scales of ROI, with same node density.For each scale, we define the number of nodes by

Nnode = αl2

where α is a density factor, and l is the side length of thesquare monitored area. In the simulation, l increasesfrom 6–30 m in steps of 3 m, and α is set to 1/3.

Figure 6 shows: (i) DSS scheme is very robust to thenetwork scale (error keeps constant when the networkscale increases) since it only schedules the activatednodes, while the error with the PSS scheme increasesquickly because of the growing number of empty de-tections; and (ii) the DSS scheme is superior to the PSSscheme, especially when the network scale is very large.

6 9 12 15 18 21 24 27 300

1

2

3

4

5

Length of the Side(meter)

Max

Err

or(m

eter

)

Impact of the Scale of the ROI

Maximum Error of PSSMaximum Error of DSS

(a) Impact of ROI scale on max error

6 9 12 15 18 21 24 27 300

0.2

0.4

0.6

0.8

1

1.2

Length of the Side(meter)

Mea

n E

rror

(met

er)

Mean Error of PSSMean Error of DSS

(b) Impact of ROI scale on mean error

Fig. 6 Impact of ROI scale on tracking error

Page 8: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

4.3 Impact of the target velocity

In the third group of simulations, we compare the per-formance of schemes under different target velocities.By setting the extent of the uniform distribution to 0.4,we vary the mean target velocity v from 0.3 to 1.9 m/s,i.e., in each set of simulations, the target velocity israndomly selected between (v − 0.2) ∼ (v + 0.2) m/s.

We depict the simulation results in Fig. 7. We cansee: (i) the DSS scheme outperforms the PSS scheme;(ii) the advantage gets more remarkable with the in-crease of the target velocity since when the targetmoves faster, the negative effect of empty detectionon tracking accuracy becomes more obvious; and (iii)The tracking error for both schemes increases linearlywith the target velocity. This is because higher targetvelocity results in larger process noise in our model,where EKF promises higher tracking error.

4.4 Impact of M in distributed sensor schedulingscheme

Figure 8 shows the impact of M to the DSS scheme.Simulation results indicate that: (i) there exists an op-timal Mo, when M equals to Mo, the tracking erroris the minimum; (ii) when M is larger than Mo, thetracking error increases with the increase of M becausethe average sampling period increases; and (iii) whenM is smaller than Mo, the tracking error reduces withthe increase of M. The reason for this phenomenon

0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.90.4

0.6

0.8

1

1.2

1.4

Averaged Target Velocity(m/s)

Max

Err

or(m

eter

)

Impact of the Target Velocity

Maximum Error of PSSMaximum Error of DSS

(a) Impact of target velocity on max error

0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.90

0.05

0.1

0.15

0.2

0.25

0.3

Averaged Target Velocity(m/s)

Mea

n E

rror

(met

er)

Mean Error of PSSMean Error of DSS

(b) Impact of target velocity on mean error

Fig. 7 Impact of target velocity on tracking error

1 3 5 10 15 20 25 30 35 40 45 500

0.2

0.4

0.6

0.8

1

M

Max

Err

or(m

eter

)

Impact of Tbackoff in DSS Scheme

Maximum Error of DSS,Node Number=12Maximum Error of DSS,Node Number=24Maximum Error of DSS,Node Number=36

(a) Impact of Tbackof f on max error in DSS Scheme

(b) Impact of Tbackof f on mean error in DSS Scheme

1 3 5 10 15 20 25 30 35 40 45 500.03

0.0350.04

0.0450.05

0.0550.06

0.0650.07

M

Mea

n E

rror

(met

er)

Mean Error of DSS,Node Number=12Mean Error of DSS,Node Number=24Mean Error of DSS,Node Number=36

Fig. 8 Impact of Tbackoff on tracking error in DSS scheme

is that the probability of more than one sensor nodestriggering their delay timer at the same time increaseswith the smaller M. Therefore, only nodes with thesmallest ID will be selected to sense the target, whichresults in the increase of accumulate error; and (iv)from the simulation results,

Mo ∝ α (8)

Equation 8 indicates that the optimal Mo is approxi-mately proportional to the node density. This is because

1 3 5 10 15 20 25 30 35 40 45 500.4

0.6

0.8

1

M

Max

Err

or(m

eter

)

Impact of Tbackoff

Maximum Error of DSSMaximum Error of Improved DSS

(a) Impact of Tbackof f on max error

(b) Impact of Tbackof f on mean error

1 3 5 10 15 20 25 30 35 40 45 500.045

0.05

0.055

0.06

M

Mea

n E

rror

(met

er)

Mean Error of DSSMean Error of Improved DSS

Fig. 9 Impact of Tbackoff on tracking error

Page 9: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

larger node density creates chances for more than onesensor nodes triggering their delay timer at the sametime, so M should be properly chosen according to thenode density when implementing the DSS scheme.

We also compare the tracking performance of theimproved DSS scheme with the DSS scheme underdifferent M. Figure 9 shows that: (i) the improved DSSscheme reduces the tracking error of the DSS scheme,especially when M is small, which also reveals theefficiency of our new collision avoidance strategy; and(ii) the tracking error of the improved DSS scheme isapproximately proportional with M.

5 Testbed experiment

A 5.2 × 5.2 m testbed has been built, as shown inFig. 10a, for testbed-scale experiments. We mount an

overhead camera right up on the ceiling of the testbedto capture the experiments or extract accurate positionsof interested objects. Each sensor node is integratedwith one main board, one PIR sensor and three activeultrasonic sensors to realize the tracking. The mainboard, which is mainly composed of Atmel128L [17]as core unit and CC2420 [18] as communication chipis used as an communication/computation module. ThePIR sensor can detect the changes of infrared energyradiation from the environment due to the movementof the target and outputs a high level signal whichis used as an external interrupt to activate the node.The ultrasonic sensor has a maximum range of 3 mand the effective angle is π/3. All the three ultrasonicsensors are connected with the main board via UARTconnection.

In our experiments, ten sensor nodes are deployedin the testbed using 3 × 3 grid form with 2.2 m

Fig. 10 Testbed experimentsfor PSS and DSS schemes

(a) Real human trajectory (b) Tracking result of PSS scheme in 10 nodes scale

(c) Tracking result of DSS scheme (d) Tracking result of PSS scheme in 40 nodes scale

Page 10: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

displacement. Specifically, two nodes are placed in themiddle of the area to provide with ±π coverage. Thethree ultrasonic sensors are fixed on the pipe with twodifferent configurations, to cover either π/2 or π anglerange. The node that covers π/2 is placed at the cornerof the field, and The nodes cover π are placed in otherpositions.

In the experiments, a moving person is consideredas the mobile target to track. Figure 10a, taken by theoverhead camera, shows the sensor node deploymentand target trajectory. We put a sequence of marks onthe testbed as the trails to follow. And the person putshis feet on the marks at one pace per second, to gen-erate identical target trails for each test. The estimatedtrajectories of PSS scheme and DSS scheme are shownin Fig. 10b and c, respectively. Compared with the truetrajectory, both estimated trajectories are very close tothe true one. The overall tracking error is about 19 cmfor DSS and 28 cm for PSS. Due to the relatively smallnetwork scale, using PSS scheme gives enough effectivemeasurements for target updates. Next we evaluatethe number of effective measurements. The PSS is 97effective measurements and the DSS is 216, which re-veals the existence of empty detection for PSS scheme.When the target moves within the field, DSS schemealways activates sensors in the vicinity of the targetand thus makes most of the measurements effective.The PSS scheme, however, emphasizes fairness amongsensors regardless of the target position and yields largenumber of empty detections.

We further test the performance of PSS schemeunder larger scale network settings. Besides the tendeployed sensors, we add another 30 virtual nodesworking around the testbed with the same deploymentand sharing the channel. The tracking result is shownin Fig. 10d. Noticeably, the tracking error of PSS isabout 1.15 m, which is very large and the correspondingnumber of effective measurements is only 27.

6 Conclusion

In this paper, we have proposed a distributed sensorscheduling (DSS) scheme for target tracking in ultra-sonic sensor network. With the scheme, the com-putation burden for each sensor can be reducedsignificantly and the scalability can be guaranteed. Ithas been demonstrated that DSS scheme outperformsthe PSS scheme, especially with large scale network.Since DSS scheme focuses on the effective measure-ment, which only considers the target presence, moreefficient sensor scheduling scheme could be designedwith the estimation/prediction techniques using the his-

torical information of the target. Furthermore, DSSscheme requires sensor nodes to keep negotiating withothers, therefore, energy efficiency may be a critical is-sue for real system implementation. In addition, track-ing and adaptive sensor scheduling for multiple targetsusing multiple modalities are also very interesting.

Acknowledgements The research is supported in part byNSFC-Guangdong joint Project grant no. U0735003; NSFC grantnos. 60736021 and 60974122; and Zhejiang Provincial NaturalScience Foundation of China under Grant No. R1100324.

References

1. Zhang F, Chen J, Li H, Sun Y, Shen X (2010) Distributedinterfering sensor scheduling for target tracking. In: Proceed-ings of IEEE ChinaCom 2010. Beijing, China

2. Chipara O, Lu C, Bailey TC, Roman G-C (2010) Reliableclinical monitoring using wireless sensor networks: experi-ence in a step-down hospital unit. In: Proceedings of ACMSenSys 2010. Zurich, Switzerland

3. Kim DH, Kim Y, Estrin D, Srivastava MB (2010) SensLoc:sensing everyday places and paths using less energy. In: Pro-ceedings of ACM SenSys 2010. Zurich, Switzerland

4. Wang C, Chen J, Sun Y (2010) Sensor network localizationusing kernel spectral regression. Wirel Commun Mob Com-put (Wiley) 10(8):1045–1054

5. Chen J, Cao K, Li K, Sun Y (2011) Distributed sensor ac-tivation algorithm for tracking with binary sensor networks.Cluster Comput (Springer) 14(1):55–64

6. Markham A, Trigoni N, Ellwood SA, Macdonald DW (2010)Revealing the hidden lives of underground animals usingmagneto-inductive tracking. In: Proceedings of ACM SenSys2010. Zurich, Switzerland

7. Kusy B, Ledeczi A, Koutsoukos X (2007) Tracking MobileNodes Using RF Doppler Shifts. In: Proceedings of ACMSenSys 2007. Sydney, Australia, pp 29–42

8. Klingbeil L, Wark T (2008) A wireless sensor network forreal-time indoor localisation and motion monitoring. In: Pro-ceedings of ACM/IEEE IPSN 2008. St. Louis, Missouri,pp 39-50

9. Li H, Miao D, Chen J, Sun Y, Shen X (2009) Networkedultrasonic sensors for target tracking: an experimental study.In: Proceedings of IEEE GlobeCom 2009. Honolulu, Hawaii

10. He Y, Chong KP (2004) Sensor scheduling for target track-ing in sensor networks. In: Proceedings of IEEE CDC 2004.Atlantis, Paradise Island, Bahamas, pp 743–748

11. Chen J, Cao K, Sun Y, Shen X (2009) Adaptive sensor ac-tivation for target tracking in wireless sensor networks. In:Proceedings of IEEE ICC 2009. Dresden, Germany

12. Huber MF, Hanebeck UD (2008) Priority list sensor schedul-ing using optimal pruning. In: Proceedings of the 11thinternational conference on information fusion. Cologne,Germany

13. Gribben J, Boukerche A, Pazzi R (2010) Scheduling for scal-able energy-efficient localization in mobile ad hoc networks.In: Proceedings of IEEE SECON 2010. Boston, MA

14. Gupta V, Chung TH, Hassibi B, Murray RM (2006) On a sto-chastic sensor selection algorithm with applications in sensorscheduling and sensor coverage. Automatica 42(2):251–260

Page 11: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

15. Xiao W, Wu J, Xie L, Dong L (2006) Sensor Scheduling forTarget Tracking in Networks of Active Sensors. Acta Auto-matica Sinica 32(6):922–928

16. Johnson DB, Maltz DA (1996) Dynamic source routing in adhoc wireless networks. In: Imielinski T, Korth H (eds) Mobilecomputing, vol 353. Kluwer Academic Publishers

17. Atmel Inc. Atmel Atmega128L Low-Power Mricrocon-troller. Available at: http://www.atmel.com/dyn/resources/proddocuments/2467S.pdf. Accessed 25 Oct 2008

18. Texas Instruments. CC2420 Single-Chip 2.4 GHz RF Trans-ceiver. Available at: http://focus.ti.com/docs/prod/folders/print/cc2420.html. Accessed 25 Oct 2008

Fan Zhang received his B.Sc degree in Control Science and En-gineering from Zhejiang University at 2009. Now he is workingtowards the Master degree under the supervise of Prof. JimingChen. His current research interests are in target tracking andactive sensor scheduling. [email protected].

Jiming Chen (M’08) received B.Sc degree and Ph.D degree bothin Control Science and Engineering from Zhejiang Universityin 2000 and 2005, respectively. He was a visiting researcher atINRIA in 2006, National University of Singapore in 2007, andUniversity of Waterloo from 2008 to 2010. Currently, he is a fullprofessor with Department of control science and engineering,and the coordinator of group of Networked Sensing and Controlin the State Key laboratory of Industrial Control Technology atZhejiang University, China. His research interests are estimationand control over sensor network, sensor and actuator network,target tracking in sensor networks, optimization in mobile sen-sor network. He currently serves associate editors for severalinternational Journals, e.g., Wireless Communication and Mobile

Computing (Wiley). He also serves as a Co-chair for Ad hoc andSensor Network Symposium, IEEE Globecom 2011, IEEEMASS 2011 Publicity Co-Chair, IEEE DCOSS 2011 Pub-licity Co-Chair, etc., and TPC member for IEEE ICDCS2010, IEEE MASS 2010, IEEE INFOCOM 2011, etc. E-mail:[email protected].

Hongbin Li received his B.E in 2005 and Ph.D in 2010 bothin Control Science and Engineering from Zhejiang Univer-sity. He is currently working at MicroStrategy China Tech-nology Center as a software performance engineer. E-mail:[email protected].

Youxian Sun received the Diploma from the Department ofChemical Engineering, Zhejiang University, China, in 1964. Hejoined the Department of Chemical Engineering, Zhejiang Uni-versity, in 1964. From1984 to1987, he was an Alexander VonHumboldt Research Fellow, and Visiting Associate Professor atUniversity of Stuttgart, Germany. He has been a full professorat Zhejiang University since 1988. In 1995, he was elevated to anAcademician of Chinese Academy of Engineering. His currentresearch interests include modeling, control and optimization ofcomplex systems, robust control design and its application. He isauthor and co-author of 450 journal and conference papers. He iscurrently the director of institute of industrial process control andnational engineering research center of industrial automation,Zhejiang University. He is President of Chinese Association ofAutomation, also served as Vice-Chairman of IFAC Pulp andPaper Committee, and Vice-President of China Instrument andControl Society. E-mail: [email protected].

Page 12: Distributed Active Sensor Scheduling for Target Tracking ...bbcr.uwaterloo.ca/~xshen/paper/2012/dasssf.pdf · mediately for other sensor nodes with pending sensing tasks. It is demonstrated

Mobile Netw Appl

Xuemin (Sherman) Shen (M’97-SM’02-F’09) received theB.Sc.(1982) degree from Dalian Maritime University (China)and the M.Sc. (1987) and Ph.D. degrees (1990) from RutgersUniversity, New Jersey (USA), all in electrical engineering.He is a Professor and University Research Chair, Departmentof Electrical and Computer Engineering, University of Water-loo, Canada. Dr. Shen’s research focuses on resource manage-ment in interconnected wireless/wired networks, UWB wirelesscommunications networks, wireless network security, wirelessbody area networks and vehicular ad hoc and sensor networks.

He is a co-author of three books, and has published morethan 400 papers and book chapters in wireless communicationsand networks, control and filtering. Dr. Shen has served asthe Technical Program Committee Chair for IEEE VTC’10,the Tutorial Chair for IEEE ICC’08, the Technical ProgramCommittee Chair for IEEE Globecom’07, the General Co-Chair for Chinacom’07 and QShine’06, the Founding Chairfor IEEE Communications Society Technical Committee onP2P Communications and Networking. He has also served asa Founding Area Editor for IEEE Transactions on WirelessCommunications; Editor-in-Chief for Peer-to-Peer Networkingand Application; Associate Editor for IEEE Transactions onVehicular Technology; Computer Networks; and ACM/WirelessNetworks, Guest Editor for IEEE JSAC, IEEE Wireless Com-munications, IEEE Communications Magazine, and ACM Mo-bile Networks and Applications, etc. Dr. Shen received theExcellent Graduate Supervision Award in 2006, and the Out-standing Performance Award in 2004 and 2008 from the Uni-versity of Waterloo, the Premier’s Research Excellence Award(PREA) in 2003 from the Province of Ontario, Canada, andthe Distinguished Performance Award in 2002 and 2007 fromthe Faculty of Engineering, University of Waterloo. Dr. Shenis a registered Professional Engineer of Ontario, Canada, anda Distinguished Lecturer of IEEE Communications Society.E-mail: [email protected].