On Topology of Sensor Networks Deployed for Tracking

Ye Zhu Anil Vikram

Department of Electrical Engineering Department of Electrical Engineering
Cleveland State University Cleveland State University
Cleveland, USA Cleveland, USA
[email protected] [email protected]

Abstract—In this paper, we study topologies of sensor net- to mean the signal transmitted from or caused by individual
works deployed for tracking multiple targets. Tracking multiple targets in the rest of the paper.
moving targets is a challenging problem. Most of previously Aggregate signals collected by sensors networks pose a
proposed tracking algorithms assume access to the signal
from individual target for tracking. Algorithms based on Blind big challenge to target-tracking solutions. Most of previous
Source Separation (BSS), a statistical signal processing tech- researches focus on simplified problem. Some researches
nique widely used to recover individual signals from mixtures assume that only one target exists in a field of interest.
of signals were proposed for tracking. BSS based tracking Some researches assume that interference from targets other
algorithm gain access to signals from individual targets by than the one of interest is negligible. Others assume that
recovering individual signals from mixture signals. The topology
of a wireless sensor network deployed for tracking with BSS- sensors can distinguish targets by tags embedded in signals
based algorithms is critical to tracking performance: (a) The or by having different targets to send signals using different
topology affects separation performance. (b) The topology channels such as using different frequency bands. Singh
determines accuracy and precision of estimation on the paths et al. [?] propose a general approach to track multiple
taken by targets. We propose cluster topologies for BSS-based targets indistinguishable by sensors. The approach is based
tracking algorithms. Guidelines on parameter selection for pro-
posed topologies are given in this paper. We evaluate proposed on binary proximity sensors that can only report whether or
cluster topologies with extensive experiments. Our empirical not there are targets in sensing area. The approach is based
experiments also show that BSS-based tracking algorithm can on simple devices and the binary proximity sensors with the
achieve comparable tracking performance in comparison with cost of limitation that it is only applicable to tracking targets
algorithms assuming access to individual signals. in smooth paths [?].
Keywords-Blind Source Separation; Topology; Tracking; We propose an approach based on Blind Source Sepa-
Sensor Networks; ration, a methodology from statistical signal processing to
recover unobserved “source” signals from a set of observed
I. I NTRODUCTION mixtures of the signals. Blind source separation models were
In this paper, we investigate topologies of sensor net- originally defined to solve cocktail party problem: The blind
work deployed for tracking multiple targets. We focus our source separation algorithms can extract one person’s voice
attention on tracking algorithms based on Blind Source signal given the mixtures of voices in a cocktail party. Blind
Separation (BSS) algorithms. source separation algorithms solve the problem based on
Tracking moving targets with wireless sensors is one of the independence between voices from different persons.
prominent applications of wireless sensor networks. De- Similarly, in the target-tracking problem, it is generally
pending on the applications, sensors with different sensing safe to assume individual signals from different targets are
modalities such as acoustic, seismic, infrared, radio, and independent. So we can use blind source separation algo-
magnetic can be deployed for tracking different type of rithms to recover individual signals from aggregate signals
targets. collected by sensors. For the cases in which individual
In general, data collected by sensors is aggregate data. In signals are dependent, blind source separation algorithms
the signal processing language, signals received by sensors based on timing structure [?] of individual signals can be
is generally mixtures of signals from individual targets. used.
For example, an acoustic sensor in a field of interest may In this paper we focus on topologies of sensor network
receive sound signals from more than one targets. Obviously to improve tracking performance. The topology of a sensor
tracking targets based on mixture signals can not be accurate network deployed for tracking is critical to tracking perfor-
results when interference from targets other than the one of mance: (a) The topology affects separation performance. For
interest is not negligible. Without loss of generality, we use better separation performance, sensors should be clustered
the term aggregate signal to mean the signal received by so that there are more sensors than the number of targets
sensor, i.e., data collected by sensors and individual signal within the sensing ranges of these sensors. It is because
BSS algorithms perform better when more numbers of deviation of error distance to measure tracking performance
mixtures are available than number of source signals. In in this paper. The error distance is defined as the nearest
general better separation performance leads to better tracking distance between a specific point in the estimated areas to
performance. (b) The number of sensors which can “hear” the actual path taken by a target as shown in Figure 2. The
a target of interest and location determine how accurate and mean and standard deviation of error distance are calculated
how precise a BSS-based tracking algorithm can estimate based on all the points in estimated areas. The mean and
the path taken by the target. standard deviation of error distance measures the accuracy
We propose cluster-based topologies for better tracking and precision of the tracking algorithm respectively. If we
performance. Our contribution in this paper can be summa- cast the evaluation of the estimation algorithm in terms of
rized as follows: evaluating a statistical estimator, the accuracy corresponds
∙ We list necessary requirements on candidate topologies. to the bias of the estimator and the precision corresponds to
∙ We propose cluster-based topologies to improve track- the variance of the estimator.
ing performance. Guidelines on selected parameters B. Requirements on Candidate Topologies
of proposed cluster topologies are given in the paper.
We evaluate proposed topologies with extensive exper- We focus on topology of low-density sensor networks sim-
iments. ply because the effect of topology on tracking performance is
∙ Our empirical experiments shows that BSS-base algo-
negligible for high-density sensor networks. In this paper, we
rithm can achieve comparable performance in compari- assume candidate topologies which should satisfy following
son with track algorithms assuming access to individual requirements:
signals with less hardware cost. ∙ Planned Deployment: The deployment used in tracking

∙ We propose metrics to evaluate performance of our targets can be classified into two categories: random
topology using target-tracking algorithms. The metrics deployment [?] and planned deployment [?], [?]. In
originate from the general metrics used to evaluate per- random deployment sensors are distributed randomly
formance of an estimator in statistics since, essentially, over the field. We eliminate random deployment from
target tracking algorithms estimate the paths based on consideration because for low-density sensor networks,
data collected from sensor networks. tracking performance of random deployment is usually
worse than tracking performance of planned deploy-
The rest of the paper is organized as follows: Section II
ment.
formally defines the problem. We describe proposed topolo-
∙ Full coverage: In planned deployment, we focus on
gies in Section III. Section IV shows results of empirical
topologies enabling sensors to cover the whole field of
experiments. We evaluate proposed topologies under various
interest. This requirement is especially important for
settings in Section V. Section VI reviews related work. We
low-density sensor networks to prevent targets disap-
conclude our paper in Section VII.
pearing from tracking. One of the reasons to eliminate
II. P ROBLEM S TATEMENT random deployment from consideration is because of
In this section, we begin with the system model and the its possibility of incomplete coverage for low-density
goal of this paper and then proceed with the requirements on sensor networks.
candidate topologies. We finish this section with description ∙ Symmetrical topology: In this paper we only consider

on application of blind source separation algorithms in target symmetrical topologies. Symmetry is desired because:
tracking. Targets can move in various directions. Symmetrical
topologies can ensure that tracking performance is
A. System Model and Goal direction-independent.
A general model of tracking targets using wireless sensor In this paper, we focus on topologies of sensor networks
networks is shown in Figure 1. Wireless sensors are de- for BSS-based tracking algorithms: Blind source separation
ployed in certain topology in a field of interest. In general, enables tracking algorithms to track multiple targets based
a wireless sensor receives signals from multiple targets. For on individual signals instead of aggregate signals. So BSS-
example, suppose acoustic sensors are deployed in Figure 1, based tracking algorithms can potentially track targets more
Sensor 𝑂1 can receive audio signals from Target 𝑠1 , 𝑠2 , and accurately and precisely and they can track targets moving
𝑠3 at the same time. So the signal received by Sensor 𝑂1 is along paths of high-frequency variation [?]. We introduce
an aggregate signal, i.e., combination of individual signals blind source separation and rationale of applying blind
transmitted by Target 𝑠1 , 𝑠2 , and 𝑠3 . source separation to the multiple targets tracking problem
The goal of this paper is to improve tracking performance below.
for tracking multiple targets with BSS algorithms. In general,
C. Application of Blind Source Separation Algorithms in
tracking algorithms output an estimated area for a path
Tracking Targets
segment as shown in Figure 2. We use mean and standard
 Figure 1: System Model Figure 2: Error Distance

1) Blind Source Separation: Blind Source Separation An example tracking algorithm based on the overlapping
(BSS) is a methodology used in statistical signal processing technique is described in [?]. A common step in BSS-
to recover unobserved source signals from a set of observed based tracking algorithms is to group 𝑚 neighboring sensors
mixtures of the signals. The separation is called blind to and apply BSS algorithms to recover individual signals
emphasize that the source signals are not observed and 𝑆1 (𝑡), 𝑆2 (𝑡), ⋅ ⋅ ⋅ , 𝑆𝑛 (𝑡) from aggregated signals received by
that the mixture is a black box to the observer. While no these 𝑚 sensors, i.e., 𝑂1 (𝑡), 𝑂2 (𝑡), ⋅ ⋅ ⋅ , 𝑂𝑚 (𝑡) in Equation
knowledge is available about the mixture, in many cases 1.
it can be safely assumed that source signals are indepen-
dent. In its simplest form [?], the blind source separation III. T OPOLOGY OF W IRELESS S ENSOR N ETWORKS FOR
model assumes 𝑛 independent signals 𝑆1 (𝑡), ⋅ ⋅ ⋅, 𝑆𝑛 (𝑡) T RACKING
and 𝑛 observations of mixture 𝑂1 (𝑡), ⋅ ⋅ ⋅, 𝑂𝑛 (𝑡) where In this section, we introduce topologies proposed for BSS-
∑𝑛
𝑂𝑖 (𝑡) = 𝑗=1 𝑎𝑖𝑗 𝑆𝑗 (𝑡). The goal of BSS is to reconstruct based tracking algorithms. Before introducing the topolo-
the source signals 𝑆𝑗 (𝑡) using only the observed data 𝑂𝑖 (𝑡), gies, we first analyze separation performance in our experi-
the assumption of independence among the signals 𝑆𝑗 (𝑡). ments and describe rationale behind the proposed topologies.
Given the observations 𝑂𝑖 (𝑡), BSS techniques estimate the
A. Separation Performance
signals 𝑆𝑗 (𝑡) by maximizing the independence between the
estimated signals. A very nice introduction to the statistical The key step in BSS-based tracking algorithms is to apply
principles behind BSS are given in [?]. The common meth- BSS algorithms to recover individual signals from aggregate
ods employed in blind source separation are minimizations signals so that tracking algorithms can have access to
of mutual information [?], maximization of nongaussianity individual signals. Obviously the performance of separating
[?], and maximization of likelihood [?]. Timing-structure out individual signals largely dictate overall tracking perfor-
based algorithms [?] can be used to recover source signals mance. To investigate the effect of topology on separation
when source signals are dependent. performance, we did a series of initial experiments with
2) Recover Individual Signals for Target-Tracking with random topologies.
Blind Source Separation Algorithms: In our tracking ap- In these initial experiments, 700 sensors are randomly
proach, blind source separation algorithms are used to re- distributed in a field of size 1.6𝑘𝑚 × 1.6𝑘𝑚. To remove
cover individual signals, i.e., source signals as described boundary issues, totally 15 targets are restricted to move
above from aggregate signals, i.e., observations as described in the field center of size 1𝑘𝑚 × 1𝑘𝑚. Sensing range of
above. Suppose acoustic sensors are deployed in the field each sensor is 250𝑚. A correlation-based metric denoted
shown in Figure 1, Sensor 𝑂1 can receive audio signals by 𝑃𝑠𝑒𝑝 is used in our experiments to measure separation
from targets 𝑆1 , 𝑆2 , and 𝑆3 and Sensor 𝑂2 can receive performance. It is calculated by taking the absolute value
audio signals from targets 𝑆2 and 𝑆4 . If we represent the of correlation between original signals and separated signal.
signal received by Sensor 𝑂𝑖 as 𝑂𝑖 (𝑡) and the signal from We use absolute value because one separated signal may
Target 𝑆𝑖 as 𝑆𝑖 (𝑡), we can have following two equations: be of different sign in comparison with the corresponding
𝑂1 (𝑡) = 𝑆1 (𝑡) + 𝑆2 (𝑡) + 𝑆3 (𝑡), 𝑂2 (𝑡) = 𝑆2 (𝑡) + 𝑆4 (𝑡). In original signal. The metric 𝑃𝑠𝑒𝑝 is within the range [0, 1].
general, for 𝑚 neighboring sensors and 𝑛 targets, we can Two topology-related metrics are used in our analysis. We
rewrite these equations in vector-matrix notation, represent the first metric as 𝐷𝑝𝑎𝑖𝑟 . It measures the average
⎛ ⎞ ⎛ ⎞ distance between each pair of sensors in a sensor group.
𝑂1 (𝑡) 𝑆1 (𝑡) The second metric measures the average distance between
⎜ 𝑂2 (𝑡) ⎟ ⎜ 𝑆2 (𝑡) ⎟ a target of interest and the center of the sensor group of
⎟ = A𝑚×𝑛 ⎜ (1)
⎜ ⎟ ⎜ ⎟
⎜ .. .. ⎟ interest when the target is moving. We represent the second
⎝ . ⎠ ⎝ . ⎠
metric as 𝐷𝑐𝑒𝑛𝑡𝑒𝑟 . It is calculated by averaging over 100 data
𝑂𝑚 (𝑡) 𝑆𝑛 (𝑡)
samples, the length of signal segments used in separation.
where A𝑚×𝑛 is called mixing matrix in the BSS literature.
Since the individual signals are independent from each other
- they come from different targets - we can use any of the Figure 3: Effect of Topology on Separation
algorithms mentioned in Section II-C1 to recover individual
signals 𝑆1 (𝑡), ⋅ ⋅ ⋅ , 𝑆𝑛 (𝑡). Given recovered individual signals, Figure 3 shows the separation performance of these initial
BSS-based tracking algorithms can use tracking techniques experiments. We present the separation performance visually
for individual signals such as overlapping sensing ranges of as a data image in Figure 3- a grid with the metrics 𝐷𝑝𝑎𝑖𝑟
sensor that can “hear” similar recovered individual signals. and 𝐷𝑐𝑒𝑛𝑡𝑒𝑟 on 𝑥-axis and 𝑦-axis respectively. The gray level
 Table II: Separation and sensor clusters are distributed evenly in a field. The
Performance vs 𝐷𝑝𝑎𝑖𝑟 proposed cluster topologies have four parameters:
Table I: Separation Per- (200𝑚 < 𝐷𝑐𝑒𝑛𝑡𝑒𝑟 <
∙ In-cluster arrangement: Within each cluster, sensors are
formance vs 𝐷𝑐𝑒𝑛𝑡𝑒𝑟 250𝑚)
arranged in regular patterns. Possible choices are any
𝐷𝑐𝑒𝑛𝑡𝑒𝑟 (m) 𝑃𝑠𝑒𝑝 𝐷𝑝𝑎𝑖𝑟 (m) 𝑃𝑠𝑒𝑝 polygon-based patterns such as well-known triangle
50 - 100 0.4841 lattice pattern, square pattern, pentagon pattern, and
50 - 70 0.681
100 - 150 0.5224
70 - 90 0.6005 hexagon pattern. Our experiments indicate tracking
150 - 200 0.5301 performance is not sensitive to patterns for in-cluster ar-
200 - 250 0.5521 90 - 110 0.5152
250 - 300 0.5132 110 - 130 0.4821 rangement. So we choose square pattern in this project
since research results on data segmentation [?], routing
[?], and storage [?] recommend the square pattern. For
same reason, we arrange clusters in the field in square
Figure 4: Example of Cluster Topology
pattern.
∙ Number of sensors per cluster 𝑛𝑐𝑙𝑠𝑡 : This parameter
specifies the number of sensors within each sensor
of each pixel in the data image represents the separation cluster. For better separation performance, the number
performance 𝑃𝑠𝑒𝑝 : A darker pixel indicates better separation of sensors per cluster should be no less than the number
performance, i.e., a larger value of 𝑃𝑠𝑒𝑝 . of targets moving in the sensing range of a cluster.
Table I summarizes separation performance in terms of Potentially all targets can move into the sensing range
the distance 𝐷𝑐𝑒𝑛𝑡𝑒𝑟 . From Figure 3 and Table I, we can of one cluster, so we choose number of sensors per
observe that the separation performance is best when the cluster close to the number of targets in the field.
distance 𝐷𝑐𝑒𝑛𝑡𝑒𝑟 is between 200m and 250m. In other words, ∙ Inter-cluster distance 𝑑𝑖𝑛𝑡𝑒𝑟 : As shown in Figure 4,
the separation performance is best when the target is away the inter-cluster distance is the distance between two
from one sensor group and still within the sensing range of centers of neighboring sensor clusters. This parameter
the sensor group. depends on 𝑁 , the total number of sensors to be
Table II shows separation performance when 𝐷𝑐𝑒𝑛𝑡𝑒𝑟 is deployed in the field and 𝑛𝑐𝑙𝑠𝑡 , the number of sensors
between 200m and 250m. We can observe from Table II that per cluster.
the separation performance is better when 𝐷𝑝𝑎𝑖𝑟 is smaller. ∙ Intra-cluster distance 𝑑𝑖𝑛𝑡𝑟𝑎 : Intra-cluster distance is
In other words, separation performance is better for sensor the distance between the center of a sensor cluster
groups with sensors closer to each other. and the furthest sensor within the same cluster. It is
These two observations are because of signal attenuation, a measure of clustering degree. To avoid neighboring
a natural consequence of signal transmission over long clusters merging, the parameter 𝑑𝑖𝑛𝑡𝑟𝑎 should be less
distances. Attenuation is a function of transmission distance. than 𝑑𝑖𝑛𝑡𝑒𝑟
2 . Our initial experiments shown in Figure
For static targets, attenuation does not affect separation 3 indicate better separation performance is achieved
performance since the distance between targets and sensors where sensors are close to each other. But it is not
does not change over time. For moving target, the distance desired to cluster sensors in a very small area because
between the target and sensors changes over time. So atten- (a) It may leave lots of uncovered spots in the field
uation becomes, from a constant for static-target cases into when sensor density is low. (b) When sensors are too
a function of time for moving-target cases. The attenuation close to each other, sensors “hear” roughly same ag-
functions for even two neighboring sensors are different. gregate signals. In this case the separation performance
So two neighboring sensors in a sensor group may “hear” can not be good. So we suggest 𝑑𝑖𝑛𝑡𝑟𝑎 to be close to
different signals from a target. 𝑑𝑖𝑛𝑡𝑒𝑟
4 . Our further experiments also support the choice
The difference causes noise in separation. Obviously, of the parameter.
when sensors are closer to each other and the target of inter- Besides considerations on separation performance, we
est is away from these sensors, the difference in attenuation propose cluster topologies because they satisfy the require-
functions of these sensors are smaller. In turn, less noise is ments listed in Section II-B: It is symmetric and it can cover
generated in separation and better separation performance the whole field. The proposed cluster topology is general:
can be achieved. Grid topology is a special case of the cluster topology when
B. Proposed Topologies 𝑛𝑐𝑙𝑠𝑡 = 1.

Based on observations made in our initial experiments, IV. E MPIRICAL E VALUATION

we propose cluster topologies for BSS-based tracking algo- We evaluate our proposed tracking algorithm using data
rithms as shown Figure 4: [?] collected from Mica2-compatible XSM motes, pro-
In proposed topologies, sensors are placed into clusters grammed using the nesC programming language and the
TinyOS operating system. The data was collected during cluster 𝑛𝑐𝑙𝑢𝑠𝑡 . For 𝑁 =128, inter-cluster distance 𝑑𝑖𝑛𝑡𝑒𝑟 =
simultaneous tracking of multiple targets as explained in [?]. 320 and 533.33 for 𝑛𝑐𝑙𝑢𝑠𝑡 =8 and 32 respectively. And for
Due to space limitation, we leave details of experiment setup 𝑁 =288, inter-cluster distance 𝑑𝑖𝑛𝑡𝑒𝑟 =228.5714 and 400
and experiment results in the companion technical report [?]. for 𝑛𝑐𝑙𝑢𝑠𝑡 =8 and 32 respectively. The number of moving
targets 𝑛𝑡𝑎𝑟𝑔𝑒𝑡𝑠 is 10 or 30 in this set of experiments. As
Table III: Performance Comparison (NA- not Applicable) shown in Figure 5 for 𝑁 =128, when 𝑛𝑐𝑙𝑢𝑠𝑡 = 8, we
Approach Mean Error Dis- Standard Devia- achieve minimum error distance for 𝑛𝑡𝑎𝑟𝑔𝑒𝑡𝑠 = 10 and for
tance (m) tion of Error Dis- 𝑛𝑡𝑎𝑟𝑔𝑒𝑡𝑠 = 30 performance is better in case of 𝑛𝑐𝑙𝑢𝑠𝑡 = 32.
tance (m) This supports our conclusion made in Section III-B. In
BSS-Based Tracking Algorithm 0.63 0.253 comparison with grid and random topology, performance
Radio Interferometric [?] 0.7 NA
increase for cluster topology is 18.75% and 23.28% for
10 and 30 moving targets when 𝑁 =128 respectively and
Table III summarizes the comparison between the BSS- when 𝑁 =288, increment is 26.12% and 27.90% for 10 and
based approach [?] and the radio interferometric approach 30 moving targets respectively. We can also observe that
[?]: The BSS-based approach can achieve comparable track- performance of cluster topology with selected parameters
ing performance. Experiments in [?], estimate the moving can be 33.62% and 32.78% better than random and grid
track by collecting frequency and phase values, where as we topologies respectively.
estimate using aggregate signal strength collected by motes.
But hardware requirement in [?] is sophisticated and costly
because of frequency and data collection by motes. Our (a) 𝑁 =128 (b) 𝑁 =288
motes are required to collect only aggregate signal strength.
Figure 5: Effect of Number of Sensors per Cluster (𝑛𝑐𝑙𝑢𝑠𝑡 )
V. P ERFORMANCE E VALUATION with 95 Percent Confidence Interval (When 𝑛𝑐𝑙𝑢𝑠𝑡 = 1,
We evaluate the performance of tracking algorithm in our Cluster Topology essentially degenerates into Grid Topol-
proposed topology with extensive simulations with Matlab. ogy.)
We assume acoustic sensors are deployed in the field of
interest for tracking purpose.
C. Effect of In-Cluster Arrangement
A. Experiment Setup
In this set of experiments, We investigate the effect of
In the following experiments, the simulated field is a In-cluster topology. We experiment on square pattern and
1.6𝑘𝑚 × 1.6𝑘𝑚 square area. The movement of targets pentagon pattern as in-cluster arrangement. Similar experi-
is restricted to a 1𝑘𝑚 × 1𝑘𝑚 center area to eliminate ment results are got from other polygons. Parameters used
boundary effects. The signals used for tracking are real bird in this set of experiments are 𝑑𝑖𝑛𝑡𝑟𝑎 =80m and 𝑛𝑡𝑎𝑟𝑔𝑒𝑡𝑠 =10.
signals downloaded from the website of Florida Museum As shown in Figure 6(a) and 6(b), we observe that tracking
of Natural History [?]. The attenuation of sound signals is performance is not sensitive to in-cluster arrangement.
according to atmospheric sound absorption model [?]. We
choose low-density sensor network with density 𝑁 =128 D. Effect of Intra-Cluster Distance (𝑑𝑖𝑛𝑡𝑟𝑎 )
and 288 sensors. The sensing range of sensors is 250m. In this set of experiments, we investigate the effect of
Details of the BSS-based tracking algorithm is described 𝑑𝑖𝑛𝑡𝑟𝑎 on tracking performance. We choose 𝑛𝑐𝑙𝑢𝑠𝑡 =8, 32 and
in [?]. In the following experiments, targets are moving vary the 𝑑𝑖𝑛𝑡𝑟𝑎 . As shown in Figure 7(a), best performance
at a speed below 0.15 meter per sample interval1 . The is achieved at 𝑑𝑖𝑛𝑡𝑟𝑎 =80m. When intra-cluster distance is
performance metrics used in our experiments is mean and very small or even close to zero, sensors within a cluster
standard deviation of error distance as described in Section observe signals from the same set of targets. So the aggregate
II-A. In all the following experiments we compare our signals received by the sensors within a cluster are close to
cluster topology with grid and random topologies. each other. In turn, it degrades the separation performance.
As shown in Figure 7(a), performance is getting better
B. Number of Sensors per Cluster (𝑛𝑐𝑙𝑢𝑠𝑡 )
with the increase of 𝑑𝑖𝑛𝑡𝑟𝑎 . But when 𝑑𝑖𝑛𝑡𝑟𝑎 > 80, error
In this set of experiments, we vary the number of sensors distance is increasing because the overlap of neighboring
per cluster 𝑛𝑐𝑙𝑢𝑠𝑡 . We choose 𝑑𝑖𝑛𝑡𝑟𝑎 =80m from the con- cluster’s sensing ranges increase and bigger overlap area
clusion drawn from Section III-B. The inter-cluster distance causes degradation in tracking performance.
depends on sensor density 𝑁 and number of sensors per
E. Effect of Number of Targets (𝑛𝑡𝑎𝑟𝑔𝑒𝑡𝑠 )
1 Due to space limitation, we leave the effect of moving speed on tracking
performance can be in our technical report [?]. Instead we focus on the In this set of experiments, we investigate the effect of
effect of topologies on tracking performance. number of targets on tracking performance. Table IV shows
 Table IV: Percentage increase in Performance of Cluster
(a) Error Distance of N=128 (b) Error Distance of N=288TopologyCompared to Grid and Random Topologies (
𝑁 =288, 𝑛𝑐𝑙𝑢𝑠𝑡 =32)
Figure 6: Effect of In-Cluster Arrangement on Tracking 𝑛𝑡𝑎𝑟𝑔𝑒𝑡𝑠 Performance Increase over Performance Increase over
Performance Grid Topology Random Topology
5 24.14 25.23
10 25.87 27.45
15 28.85 29.93
(a) Mean Error Distance (b) Standard Deviation of Error Distance
20 30.22 30.94
25 30.47 31.28
Figure 7: Effect of Intra Cluster Distance (𝑑𝑖𝑛𝑡𝑟𝑎 ) on Track- 30 32.78 33.62
35 33.92 35.13
ing Performance with 95 Percent Confidence Interval 40 35.02 37.18

the typical performance increase of cluster topology over triangular pattern which is optimal in terms of number
grid and random topologies. The experiment parameters are of circles needed to entirely cover a given area.
as follows: density 𝑁 =288, 𝑛𝑐𝑙𝑢𝑠𝑡 =32, and 𝑑𝑖𝑛𝑡𝑟𝑎 =80m. ∙ Localization: Zou and chakrabarty [?] proposed a sen-
More experimental results can be found in the companion sor deployment for target localization based on virtual
technical report [?]. We can observe that percentage of forces. Chan et.el [?] propose to use clustering for
increase can achieve 37.18% and 35.02% for 40 targets over localization in sensor network.
random topologies and grid topologies respectively. Table ∙ Communication Cost: Kasetkasem and Varsheny [?]
IV shows that performance increase becomes larger when studied placement of sensors to reduce communication
the number of targets increases. It is mainly because better cost.
separation performance can be achieved for cluster topology.
VII. C ONCLUSION
VI. R ELATED W ORK We propose cluster topologies to improve tracking perfor-
Tracking multiple targets in a field is a challenging prob- mance of BSS-based tracking algorithms. A set of guidelines
lem compared with tracking single target. Various advanced on parameter selection for proposed topologies are given in
techniques have been applied to solve the problem. The ex- the paper. We evaluate proposed topologies with extensive
amples are signal processing techniques such as wavelet [?], experiments. The proposed topology can achieve more than
statistical techniques such as principle component analysis 35 percent improvement in tracking performance over grid
[?]. A string of researches on tracking targets with wireless and random topologies. Our empirical experiments show
sensor networks are based on binary proximity sensors that BSS-based tracking algorithm can achieve comparable
which can only report whether there are targets within tracking performance compared with algorithms assuming
sensing areas. The initial work [?] on binary proximity access to individual signals.
sensors focuses on tracking single target. Singh et al. [?] R EFERENCES
extended the approach to track multiple indistinguishable
targets by applying particle filtering algorithms. Approaches
based on binary proximity sensors have two obvious ad-
vantages: (a) The sensors are very simple since they only
report binary information. (b) The approaches are robust
since interference from other targets are essentially filtered
out by an equivalent low-passed filter [?]. The cost of using
these simple devices is loss of information that is helpful
to accurately track targets due to the filtering effect. So,
approaches based on binary proximity sensors can not track
target in a path with high-frequency variations [?].
Topologies of sensor networks have been studied for
different purposes:
∙ Coverage: In [?], proposed optimal deployment to
achieve full coverage. Different patterns like polygon,
diamond and square are proposed for sensor deploy-
ment topologies. To ensure full coverage and connec-
tivity are proposed in [?]. In [?], Kershner proposed a