Multi-target Data Aggregation and Tracking in Wireless ... - CiteSeerX

JOURNAL OF NETWORKS, VOL. 3, NO. 1, JANUARY 2008

1

Multi-target Data Aggregation and Tracking in Wireless Sensor Networks Maarten Ditzel, Caspar Lageweg, Johan Janssen, Arne Theil TNO Defence, Security and Safety, The Hague, The Netherlands email: [email protected]

Abstract— This paper presents the results of a study on the effects of data aggregation for multi-target tracking in wireless sensor networks. Wireless sensor networks are normally limited in communication bandwidth. The nodes implementing the wireless sensor network are themselves limited in computing power and usually have a limited battery life. These observations are recognized and combined to come to efficient target tracking approaches. The main question to be answered is how to accurately track multiple targets crossing an area observed by a wireless sensor network, while limiting the amount of network traffic. Limiting the amount of network traffic reduces the required bandwidth and reduces the required energy. Various computing power aware data aggregation strategies are researched. They have been tested in a simulation environment and compared with each other. The results of the simulations clearly show the benefit of the new data aggregation strategies in terms of energy consumption and tracking accuracy. Index Terms— Wireless sensor networks, multi-target tracking, distributed tracking, data aggregation.

I. I NTRODUCTION Wireless sensor networks [1], [2] have experienced increasing attention in academic, industrial and military environments over the past few years. These networks promise an easy-to-deploy, easy-to-use and moreover, low-cost means to remotely monitor environments. Furthermore, sensing accuracy can be improved significantly by processing and combining collected data within the network itself. Finally, the network can be made robust to the failure of individual nodes, which ensures that the lifetime and proper operation of the network is not limited to the lifetime of one node in particular. Applications, either envisioned or already realized, are generally related to the remote monitoring of a, possibly inaccessible or hostile, environment. Examples are an aqueous surveillance system for a drinking water reservoir [3], a wildlife habitat observation system [4], and a network to monitor the behavior of glaciers [5]. Additionally, numerous military applications are envisioned such as a battlefield data collection network as described in [6]. In all the aforementioned applications, the network consists of tens to thousands of tiny devices (e.g., see Figure 1). Each device carries one or more sensors and has limited signal processing and communication capabilities. Usually, the devices are powered by batteries and can thus only operate for a limited time period. Key to implementing a network with such devices is that energy, computing power and communication bandwidth are scarce. © 2008 ACADEMY PUBLISHER

2.5 cm

Figure 1. Example of a wireless sensor node: the TNOdes (courtesy of SOWNet Technologies B.V.).

This paper presents the results of a study on the tracking of several objects in a sensor network simultaneously, focusing on the trade-offs between the amount of communication in the network and tracking accuracy. A reduction of the amount of messages sent can be achieved by utilizing local processing and sensor data aggregation within the network. During aggregation, local sensor readings are combined to reduce the communication load of the network. As a consequence, a central track algorithm receives potentially less information, possibly degrading the accuracy of the tracks. The objective of this study is to reduce the amount of messages sent within the network, while maintaining high tracking accuracy. In our simulations we assume a semi-regular grid-based network through which several objects of interest (socalled targets) travel. As we are primarily interested in the effects of the aggregation strategies on the tracking performance, we assume ideal communication links. Faulty sensor readings are included in the simulations, as these might negatively influence the communication reduction and the tracking accuracy. We investigated four strategies for the aggregation and transfer of sensor data to the central tracker. To evaluate the aggregation strategies, two basic metrics are utilized: the total amount of messages sent in the network (both broadcasts and unicasts), and the quality of the position estimates expressed as the RMS difference of the estimated track and the original path of the observed target. The remainder of this paper is organized as follows. Section II shortly discusses related work. Section III describes the characteristics of the simulation environment. The sensor data aggregation strategies are described in Section IV. Moreover, it describes the algorithm used in the simulations for target position estimation. The track algorithm is explained in Section V. Then, Section VI

2


20 m

summarizes and discusses the simulation results. Finally, Section VII states the conclusions.

51

Recent research efforts are dedicated towards operation of the network, focussing on energy efficient and robust communication schemes, reconfigurability, security, etc. [7], [8], [9]. Relatively less attention is payed to the actual goal of the networks: the collection and delivery of interesting information, extracted from data that is gathered by one or more sensors. General object localization and tracking in sensor networks is actively researched and addressed in several papers, for example [10], [11], [12], [13], [14]. The authors focus on different aspects of tracking in sensor networks, such as real-time implementation aspects, sensor query systems, classification issues and node activation strategies. A survey on node localization and tracking techniques can be found in [15]. However, data aggregation methods for target tracking and its impact on the accuracy as such, are not covered in detail. Also more recent publications, such as [16] and [17] do not address these issues. Initial results on the aggregation strategies were already presented in [18]. In this paper this work is extended in several areas1 . First, the aggregation strategies are compared for multi-target tracking. Moreover, the effects of noise (which manifests itself as false contacts) on the tracking accuracy is investigated. Finally, a central track algorithm is added to associate the individual measurements with tracks and estimate the targets’ positions and velocities. Also, the tracker ideally separates ‘true’ targets from the false contacts (introduced by the measurement noise). III. S IMULATION E NVIRONMENT Wireless sensor networks typically consist of a large number of individual nodes with limited communication and observation ranges. An object of interest (target) that passes through such a network is only ‘visible’ to nodes within whose sensor range it is located. These nodes will then attempt to send the observation data in the form of messages to a collection point, commonly referred to as the sink. Given the limited communication range of sensor nodes, sending a message from a node to the sink typically results in a series of hops through the network. Each of these hops results in the consumption of limited (battery) energy, which eventually results in failures within the network as nodes completely run out of energy (particularly in the vicinity of the sink where messages converge). For our simulations, we utilized an event driven simulator that was developed using the OMNeT++ simulation framework [19]. The simulated wireless sensor network consists of a grid of 51×51 sensor nodes. It is assumed 1 With

kind permission of Springer Science and Business Media.

© 2008 ACADEMY PUBLISHER

50

49

5

4

100 m

II. R ELATED W ORK

20 m

95%

rc 3

2

1 1

2

3

4

5

49

50

51

Figure 2. Topology of the sensor network. The nodes are placed in a semi-regular grid, such that 95% of the nodes are within 10% range of the grid points.

that several targets are traveling through the area monitored by the wireless sensor network. The observations are obscured by noise, which manifests itself as false contacts. All nodes are assumed to communicate error free. They can either communicate with each other through broadcasts (reaching all direct neighbors), or through unicast (peer-to-peer) communication, using idealized shortest path routing to reach the sink. A. Sensor Node Model The sensor node model is based on the TNOdes wireless sensor node (see Figure 1). Each node is equipped with a radio to communicate its observations. The communication range is limited. In our simulations, we assume that the sensor nodes can only communicate with their nearest neighbors. In addition, the sensor nodes can detect the range to a target using a simple sensor such as for example the radar described in [20]. The sensor update rate is 2 Hz, a common update interval for unattended ground sensors enabling the detection of pedestrians and vehicles at various speeds. The sensor range of each sensor node is such that a target travelling within the network will be observed by multiple nodes simultaneously (a dense network). All nodes have a local processing unit to process the sensor data. B. Network Topology In our simulations the wireless sensor network consists of a 51×51 grid of sensor nodes, placed 100 m apart, as indicated in Figure 2. The communication range rc of the network is chosen such that nodes can only communicate with their direct neighbors. The sink is located at the center of the network. Consequently, all nodes can reach the sink with at most 25 hops. It is assumed that the


½

¼

3

TABLE I. S UMMARY OF THE AGGREGATION STRATEGIES 1 TO 4.

¼

Strategy Strategy Strategy Strategy Strategy

30°

target path

30°

IV. S ENSOR DATA AGGREGATION

30°

Figure 3. Target trajectories. Target enter the network at four location ranges (indicated by the solid lines) at a randomly chosen angle between -15 and 15 degrees.

positions of the individual nodes are known by the nodes themselves. The sensor nodes are deployed in such a manner that the area that needs to be observed is uniformly covered. The coverage γ of the network, is defined by N πrs 2 o2 , N − 1 d + 2rs

The main goal of the research presented in this paper is to investigate various approaches to reduce the number of required messages, while achieving a certain track accuracy. We distinguish two message types: • Local messages: The messages are broadcasted (only reaching all direct neighbors); these messages do not propagate through the network. • Global messages: These messages are peer-to-peer messages and are sent to the sink using a hopping algorithm (shortest path). The data content of both messages is the same and can be described with the tuple {~x, q(t)}, where ~x describes the location and q the quality of the observation.

(1)

where N is the number of nodes, rs the sensing range and d the inter-node distance. The coverage in the simulated scenarios is approximately 7, i.e., on average 7 nodes are sensing a target travelling through the network. In realistic deployment scenarios, a perfect regular grid is unlikely. Therefore a Gaussian probability density function with zero mean is used to model the deviations in the positioning of the sensor nodes. The standard deviation is set to 5% of the inter-node distance (i.e., 5 m), resulting in roughly 95% of the nodes placed within 10% of the inter-node distance around the grid points.

A. Aggregation Strategies In order to investigate the amount of messages sent out by nodes in the network we have conducted a series of simulations. Each simulation uses a different strategy to send sensor data from observing nodes to the sink. •

Strategy 1: Reference. The first strategy does not use any local processing or data aggregation algorithm. During each time step, all nodes that observe the target report their findings to the sink using message hopping.

•

Strategy 2: Differential messaging. The second strategy uses local processing. The first time step where a node observes a target, it reports its observation to the sink. From this time step forward, it is assumed by the sink that the node can observe that target. When the previously sensed target moves through the network, the observed quality changes. When the quality difference exceeds a given threshold ∆T , the node informs the sink of this change. Also when targets move out of sensor range, the node reports it has lost the target. In other words, it only reports differential information (mutations in observations).

•

Strategy 3: Local aggregation. The third strategy uses data aggregation. During

C. Target Model In our simulations, targets can enter the sensor grid at four location ranges as shown in Figure 3. Within these ranges, the exact entry location is chosen randomly. Also, the entry angle is selected randomly between -15 and 15 degrees. In addition, on its way through the sensor grid the direction changes randomly with a small amount. The same holds for the velocity of the target, it changes randomly with an average speed of around 4 km/h, which corresponds to a human’s walking speed. In real-world scenarios, sensors are not ideal. Typically, the measurements contain noise. In our simulations, noise © 2008 ACADEMY PUBLISHER

Local aggregation – – + +

manifests itself as false contacts, i.e., the sensor detects a target that is actually not there. False contacts appear randomly in time in the network according to a Poisson process with a given rate. At each appearance of a false contact, its location is chosen randomly using a uniform probability density function.

30°

γ = n√

1 2 3 4

Differential messaging – + – +

4


Strategy 4: Local aggregation and differential messaging. The fourth strategy combines the target position estimation described for simulation 3 with the differential messaging utilized in simulation 2. Nodes that have better observation quality than what is reported to them by others (their neighbors), send a single message that informs the sink of the estimated position. As the target moves through the network, the nodes continuously estimate targets’ positions. If the distance between the last reported position and the last calculated position exceeds a given threshold, the new position is sent to the sink. Once a node no longer has the best quality observation, a single message is sent to the sink to report this change. The strategies are summarized in Table I. It should be noted that strategies 1 and 3 are in fact special cases of strategies 2 and 4, respectively, by setting their thresholds for the differential messaging to zero. Then, as all changes are reported to the sink, the only difference is the communication overhead incurred by the ‘contact lost’ messages. The next section describes the position estimation algorithm based on the collected sensor data. The algorithm is executed centrally at the sink in simulations 1 and 2. In the simulations 3 and 4 it is run locally at the nodes.

150 error [m]

each time step, all nodes that observe the target broadcast their findings to their neighbors. As part of this message they send an indication of the quality of the observation (i.e., the normalized strength of their sensor signal, see Section IV-B). Periodically, each node that observed the target compares the quality of its own observation with those reported in the broadcasts that it has received. If the quality of the node’s own observation is better than those reported by others, it estimates the target’s position by weighing each observing node’s position with the quality of their observations. The estimates are reported to the sink.

100 50 0 400 300

400 300

200 200

100 y position [m]

100

0

0

x position [m]

•

B. Position Estimation Once a target comes within sensor range of node i, its distance to the node is measured. We express the quality of this observation as qi ∈ [0, 1], where qi = 0 implies that the target is out of sensor range and qi = 1 that the target is at the same position as the sensor node. Each node i observing the target is assumed to broadcast {~xi , qi (t)}, where ~xi is the position of the node, and qi (t) the quality of the observation at time t, which is calculated using |~xi − ~xo (t)| , (2) qi (t) = 1 − rs

where ~xo (t) is the time dependent position of the observed target, and rs the maximum range of the node’s sensor. If the quality of a node’s own observation is higher than that reported by others, the node calculates the sum of observation qualities (its own and those reported by © 2008 ACADEMY PUBLISHER

Figure 4. Error between a target’s ‘true’ position and its estimated value in a regular 7×7 sensor grid (the error is plotted only for a quarter of the network). The dots depict the locations of the sensor nodes.

others) which it uses to determine the weight wi (t) of each observation as qi (t) (3) wi (t) = P n∈N qn (t)

where N represents the set observing neighbors and the node itself. Given the weights wi (t), we calculate the ˆo (t) as the weighted sum of target’s estimated position ~x the observing nodes’ positions X ˆo (t) = wn (t)~xn . (4) ~x n∈N

To analyze the estimates performance, the error between a target’s ‘true’ position and the estimated value is plotted in Figure 4. The error is plotted for a regular 7×7 grid. For clarity, the error is plotted for only a quarter of the network. V. T RACKING IN W IRELESS S ENSOR N ETWORKS In the context of this paper, tracking is defined as the process of following the movements and establishing the location of an object or person. The quality of the observations in combination with the fitness of the motion model that is applied in the track algorithm determine the overall performance of our wireless sensor network. Especially in the context of multiple targets and false contacts, a sophisticated track algorithm greatly improves the performance. The track algorithm is assumed to run at the sink node. It is fed with the observations (the tuples {~x, q(t)}) it receives from the wireless sensor network. The tracker determines whether a contact (observation) can be associated with a track, representing the estimated trajectory of a target through the network. Ideally, only false contacts do not pass the association criterion. A. Kalman Filter To determine a target’s trajectory a Kalman filter is used. The Kalman filter is a recursive filter, which estimates the state of a dynamic system from a series of noisy


5

measurements. The Kalman filter exploits the dynamics of the target to remove the effects of noise and to get a good estimate of the location of the target. Apart from the original publication [21], details of Kalman filtering can be found in numerous ‘standard’ textbooks on digital signal processing. B. Gating A track is constituted from a collection of contacts that ideally belong to the same target. The state covariance matrix that is estimated by the Kalman filter is used to determine with which target/track a new measurement can be associated. To accomplish this, the Kalman filter predicts the locations of the targets for each observation. Around the predicted location a so-called track gate is formed. The size and the shape of the track gate must be chosen in such a way that unlikely measurements are precluded from the contact-track association process, while the probability of a measurement lying outside the track gate is small. New observations within the track gate are associated with the corresponding track/target. Observations that cannot be associated with a track are considered to be originating from a new target. A track is discarded (track loss) when the track has not had updates for a certain period of time. C. Multi Stage Contact-Track Association False contacts can result in false tracks. To reduce the number of false tracks, multi-stage contact-track association is applied. The technique labels tracks in three categories: potential, tentative and confirmed. A contact that does not associate with any of the available tracks is promoted to a potential track. If suitable contacts in the vicinity of this contact are generated within a predefined time interval, the track status transits to tentative. Whenever the tentative track consists of a certain number of contacts (e.g., four, five) it gets the confirmed status. Figure 5 presents the state transition diagram. The indicated state transition parameters (Npt , Ntc , Tpl , Ttl and Tcl ) are properly adjusted to the specific sensors used and the behavior of the expected targets. Npt and Ntc indicate the number of contacts necessary to promote to the next state. Tpl , Ttl and Tcl set the expiration times for each state before the track is considered lost. VI. R ESULTS The communication that occurs during the experiments can be divided in three categories of interest. First, one can count the total amount of messages that reach the sink and utilize this as an indication of the energy consumption of the nodes along the path from source to sink. Second, one can count the total amount of messages that are sent from one node to another (unicast and broadcast) as an indication of total energy consumption. Third, one can examine the fraction of broadcasted messages separately as broadcasts typically result in higher energy © 2008 ACADEMY PUBLISHER

Figure 5. State transition diagram of the tracker.

consumption due to the fact that more nodes are listening2 . We have analyzed these three categories for the four data aggregation strategies described previously and summarized the results in Table II and Figure 6. Examining Table II one can observe the following. First, the total amount of messages in the simulations is equal to or slightly larger than the number of messages that arrived to the sink multiplied with a factor of around 13. This factor is roughly determined by the average hopping distance to the sink. When increasing or decreasing the size of the network, one can expect to see this factor increase and decrease proportionally as the average distance of the nodes to the sink varies. Second, examining the data from the simulations of strategies 3 and 4 (local aggregation), while taking the results of strategy 1 as a basis for comparison, one observes that each node that previously sent their observations to the sink now broadcasts its observations to their neighbors. When changing the size of the network, the number of broadcasts will remain constant if the average number of nodes that observe the target remains unchanged (this number is given by the coverage γ of the network). As previously stated, in our simulation scenarios on average 7 nodes observe the target simultaneously. Apart from the cost in terms of transmitted messages, the influence of the different aggregation strategies is analyzed. The results of the track algorithm are depicted in Figure 7 for most of the simulations. Figure 7(a) depicts the actual route of the target and Figure 7(b) to Figure 7(f) plot the tracks calculated in simulations 1 to 4, respectively. For clarity only a small area of the actual sensor grid is plotted. From these figures, performance metrics can be can be derived for each data aggregation strategy. They are calculated as the RMS difference between the estimated 2 In a typical sensor node transmitting and receiving require approximately the same energy. For example, in the TNOdes (see Figure 1, transmission and reception draw 16 mA and 12 mA, respectively.

6


2

2

10 Exp. #1 Exp. #2, ∆T=0

messages/observation


10

Exp. #2, ∆T=2

1

10

Exp. #2, ∆T=5 Exp. #2, ∆T=10

0

10

−1

10

0

Exp. #1 Exp. #3 1

10

0

10

−1

5

15 10 hopping distance

10

25

20

5

0

(a) Aggregation strategy 2

(b) Aggregation strategy 3

2

2

10 Exp. #1 Exp. #4, ∆T=0



10

Exp. #4, ∆T=2

1

10

Exp. #4, ∆T=5

0

10

−1

10

25

20


0

Exp. #1 Exp. #2, ∆T=5 Exp. #3 Exp. #4, ∆ =5

1

10

T

0

10

−1

5


10

25

20

5

0

(c) Aggregation strategy 4


20

25

(d) Summary

Figure 6. Total number of messages per observation (unicast and broadcast) as function of the number of hops to the sink (hopping distance). Aggregation strategies 2 to 4 are compared with the reference strategy. TABLE II. C OMMUNICATION METRICS FOR DATA AGGREGATION STRATEGIES 1 TO 4. VALUES INDICATE THE AVERAGE NUMBER OF MESSAGES PER OBSERVATION .

Strategy 1 - Reference 2 - Differential messaging 2 - Differential messaging 2 - Differential messaging 2 - Differential messaging 3 - Local aggregation 4 - Local aggregation and differential messaging 4 - Local aggregation and differential messaging 4 - Local aggregation and differential messaging

∆T 0% 2% 5% 10% 0% 2% 5%

sink arrivals 8.92 9.87 2.70 2.21 2.04 1.29 2.39 1.49 1.43

paths and the actual route of the target. Figure 8 shows the cost-performance tradeoff for each strategy. The cost is expressed as the total number of messages sent per observation during the simulations. Analyzing the results, several observations can be made. First, and foremost, local processing can greatly reduce the communication load of the network, either using differential messaging, local aggregation, or both. Second, the track accuracy alone is not significantly degraded when introducing aggregation in combination with a central tracker. Finally, as differential messaging has a relatively high communication penalty for false contacts (it sends at least two messages: one on first contact, and one when the contact is lost), it looses its advantage over local aggregation only, as previously observed in [18]. VII. C ONCLUSIONS AND F UTURE W ORK This paper discusses the results of a study on the effects of data aggregation for target tracking in wireless sensor networks. Various novel aggregation strategies © 2008 ACADEMY PUBLISHER

Unicasts 125.96 141.86 42.94 36.21 33.90 18.78 37.23 24.83 23.97

Broadcasts 0.00 0.00 0.00 0.00 0.00 8.92 8.92 8.92 8.92

Total 125.96 141.86 42.94 36.21 33.90 27.71 46.15 33.75 32.90

Broadcast ratio 0.0 0.0 0.0 0.0 0.0 32.2 19.3 26.4 27.1

have been implemented and analyzed, using an event driven simulation environment. For all simulations we used a target position estimation algorithm described in this paper, either run centrally at the sink, or locally at the nodes. The estimation positions were fed to a track algorithm to estimate the trajectories of several targets travelling through the network. Furthermore, the track algorithm filters out false contacts, which may degrade the quality of the estimates. The amount of communication for each strategy is analyzed as it largely accounts for the energy consumption in wireless sensor networks. It is compared with the track accuracy of the different data aggregation strategies. The results from the simulations clearly show a relevant trade-off between the amount of communication and the performance of the track algorithm. Moreover, local aggregation without differential messaging (as used in the third simulation) has the best cost-performance trade-off in noisy environments. Current activities are focussed on implementing the var-


7

3000

400 2000

300

y position [m]

y position [m]

1000

0

200

100

−1000

0 −2000

−100 −3000 −3000

−1000

−2000

0 x position [m]

2000

1000

400

3000

400

400

300

300

200

100

−100

−100 700 600 x position [m]

800

900

400

400

400

300

300

200

100

800

900

100

0

−100

−100 700 600 x position [m]

800

700 600 x position [m]

200

0

500

500

(d) Strategy 3 – Local aggregation

y position [m]

y position [m]

(c) Strategy 2 – Differential messaging, ∆T = 5

400

900

100

0

500

800

200

0

400


(b) Strategy 1 – Reference

y position [m]

y position [m]

(a) Actual route of the target

500

900

(e) Strategy 4 – Local aggregation and differential messaging, ∆T = 2

400

500


800

900

(f) Strategy 4 – Local aggregation and differential messaging, ∆T = 5

Figure 7. Target tracking results. The dashed lines represent the actual route of the target. The circles depict the locations of the sensor nodes, whereas the crosses indicate contacts (both true and false).


8


25

2(10)

estimate error [m]

20

15

10

2(5)

3 4(5) 2(2) 4(2) 4(0)

1 2(0)

5

0 1 10

2

10 messages/observation

3

10

Figure 8. Cost-performance trade-offs: estimation error versus the total number of messages per simulation. The index x(y) denotes the xth strategy, with threshold y (if applicable).

ious strategies in a real-world test-bed. Moreover, we are investigating the possibilities of running parts of the track algorithm distributedly at the nodes, instead of centrally. A promising candidate for a distributed implementation is early false contact suppression, which may reduce the communication load of the network even further. R EFERENCES [1] D. Estrin, R. Govindan, J. Heidemann, and S. Kumar, “Next century challenges: scalable coordination in sensor networks,” in ACM/IEEE international conference on Mobile computing and networking. ACM Press, 1999, pp. 263–270. [2] G. J. Pottie and W. J. Kaiser, “Wireless integrated network sensors,” Commun. ACM, vol. 43, no. 5, pp. 51–58, 2000. [3] X. Yang, K. G. Ong, W. R. Dreschel, K. Zeng, C. S. Mungle, and C. A. Grimes, “Design of a wireless sensor network for long-term, in-situ monitoring of an aqueous environment,” Sensors, vol. 2, no. 11, pp. 436–472, November 2002. [4] A. Mainwaring, D. Culler, J. Polastre, R. Szewczyk, and J. Anderson, “Wireless sensor networks for habitat monitoring,” in International workshop on Wireless sensor networks and applications. ACM Press, 2002, pp. 88–97. [5] K. Martinez, J. K. Hart, and R. Ong, “Environmental sensor networks,” Computer, vol. 37, no. 8, pp. 50–56, August 2004. [6] F. Ye, H. Luo, J. Cheng, S. Lu, and L. Zhang, “A two-tier data dissemination model for large-scale wireless sensor networks,” in International conference on Mobile computing and networking. ACM Press, 2002, pp. 148– 159. [7] T. van Dam and K. Langendoen, “An adaptive energyefficient MAC protocol for wireless sensor networks,” in International conference on Embedded networked sensor systems. ACM Press, 2003, pp. 171–180. [8] W. Ye, J. Heidemann, and D. Estrin, “An energy efficient MAC protocol for wireless sensor networks,” in Conference of the IEEE Computer and Communications Societies (INFOCOM), vol. 3, June 2002, pp. 1567–1576. [9] J. Kulik, W. Heinzelman, and H. Balakrishnan, “Negotiation-based protocols for disseminating information in wireless sensor networks,” Wirel. Netw., vol. 8, no. 2/3, pp. 169–185, 2002.


[10] A. Arora, P. Dutta, S. Bapat, V. Kulathumani, et al., “A line in the sand: A wireless sensor network for target detection, classification, and tracking,” Computer Networks, Special Issue on Military Communications Systems and Technologies, vol. 46, no. 5, pp. 605–634, July 2004. [11] B. Horling, R. Vincent, R. Mailler, J. Shen, R. Becker, K. Rawlins, and V. Lesser, “Distributed sensor network for real time tracking,” in International conference on Autonomous agents. New York, NY, USA: ACM Press, 2001, pp. 417–424. [12] F. Zhao, J. Shin, and J. Reich, “Information-driven dynamic sensor collaboration for target tracking,” IEEE Signal Processing Magazine, vol. 19, no. 2, pp. 61–72, March 2002. [13] D. Li, K. Wong, Y. H. Hu, and A. Sayeed, “Detection, classification and tracking of targets in distributed sensor networks,” IEEE Signal Processing Magazine, vol. 19, no. 2, pp. 17–29, 2002. [14] S. Pattem, S. Poduri, and B. Krishnamachari, “Energyquality tradeoffs for target tracking in wireless sensor networks,” in International Symposium on Aerospace/Defense sensing Simulation and Controls, Aerosense, April 2003. [15] J. Hightower and G. Borriella, “Location systems for ubiquitous computing,” IEEE Computer, vol. 34, no. 8, pp. 57–66, August 2001. [16] B. Kusy, G. Balogh, J. Sallai, A. Ledeczi, and M. Maroti, “Intrack: High precision tracking of mobile sensor nodes,” in European conference on Wireless Sensor Networks, ser. Lecture Notes in Computer Science, vol. 4373. Springer, 2007, pp. 51–66. [17] V. Kulathumani, M. Demirbas, A. Arora, and M. Sridharan, “Trail: A distance sensitive wireless sensor network service for distributed object tracking,” in European conference on Wireless Sensor Networks, ser. Lecture Notes in Computer Science, vol. 4373. Springer, 2007, pp. 83–100. [18] C. Lageweg, J. Janssen, and M. Ditzel, “Data aggregation for target tracking in wireless sensor networks,” in European Conference on Smart Sensing and Context, ser. Lecture Notes in Computer Science, vol. 4272. Springer, 2006, pp. 15–24. [19] A. Varga, “Omnet++: Objective modular network testbed in C++,” http://www.omnetpp.org. [20] M. Ditzel and F. Elferink, “Low-power radar for wireless sensor networks,” in European Radar Conference, September 2006. [21] R. E. Kalman, “A new approach to linear filtering and prediction problems,” Transactions of the ASME–Journal of Basic Engineering, vol. 82, no. Series D, pp. 35–45, 1960.

Maarten Ditzel received the MSc degree with honors in electrical engineering at Delft University of Technology in 1998. In 2004, he received his PhD degree from the micro-electronics department of the same university, where he worked in the Ubiquitous Communications program. In 2003 he joined TNO’s Physics and Electronics Laboratory in The Hague. Today he is working as a senior research scientist at TNO Defence, Security and Safety. His current research interests are with distributed embedded sensing and processing. Johan Janssen studied Electrical Engineering at the Delft University of Technology, where he joined the Information Theory group. In 1993 he graduated with honors. Subsequently, he has been a PhD student at the Computer


Engineering group at the department of Electrical Engineering of the Delft University of Technology. In 2001 he received his PhD degree. Early 1998 he joined TNO’s Physics and Electronics Laboratory in the Hague, where he was involved in embedded system design and research towards networked intelligent devices. Today, his work at TNO Defence, Security and Safety concentrates on project management and development of new technologies and applications. Caspar Lageweg received the MSc degree in electrical engineering from Delft University of Technology. After completing his degree, he worked for Hewlett-Packard laboratories in Bristol, United Kingdom until 2001. In 2004 he received his PhD from Delft University of Technology.


9

Since 2005 he has been working at TNO Defence, Security and Safety, where he is involved in distributed system design. Arne Theil received his MSc degree from the Pattern Recognition Group of the Physics department of the Delft University of Technology. Since August 1985 he works at TNO Defence, Security and Safety in The Hague in the radar section of the business unit Observation Systems. His work concerns antenna signal processing and contact extraction for array radars, Doppler polarimetry, tracking and sensor fusion. He has co-authored the popular software program CARPET (Computer Aided Radar Performance Evaluation Tool, Artech House 1982) for radar performance assessment.