of 12
7/23/2019 07177096(1)
1/12
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 63, NO. 22, NOVEMBER 15, 2015 6049
Compressive Sensing Based Probabilistic
Sensor Management for Target Tracking in
Wireless Sensor NetworksYujiao Zheng, Student Member, IEEE, Nianxia Cao, Student Member, IEEE, Thakshila Wimalajeewa, Member, IEEE,
and Pramod K. Varshney, Fellow, IEEE
AbstractIn this paper, we consider the problem of sensor man-agement for target tracking in a wireless sensor network (WSN).To determine the set of sensors with the most informative data, wedevelop a probabilistic sensor management scheme based on theconceptsdeveloped in compressive sensing. In the proposed scheme
where each sensor transmits its observation with a certain prob-ability via a coherent multiple access channel (MAC), the obser-vation vector received at the fusion center becomes a compressedversion of the original observations. In this framework, the sensor
management problem can be cast as the problem of finding theprobability of transmission at each node so that a given perfor-mance metric is optimized. Our goal is to determine the optimalvalues of the probabilities of transmission so that the trace of theFisher information matrix (FIM) is maximized at any given timeinstant with a constraint on the available energy. We consider twocases, where the fusion center has i) complete information and ii)
only partial information, regarding the sensor transmissions. Theexpression for FIM is derived for both casesand the optimal valuesof the probabilities of transmission are found accordingly. Withnonidentical probabilities, we obtain the results numerically while
under the assumption that each sensor transmits with equal prob-ability, we obtain the optimal values analytically. We provide nu-
merical results to illustrate the performance of the proposed prob-abilistic sensor management scheme.
Index TermsCompressive sensing, sensor management, targettracking, wireless sensor networks.
I. INTRODUCTION
A typical wireless sensor network (WSN) is composed ofa large number of densely deployed sensors, where sen-sors are assumed to be tiny, battery-powered devices with lim-
ited signal processing capabilities. When programmed and net-
worked properly, WSNs are very useful in many application
Manuscript received August 01, 2014; revised March 15, 2015 and July 13,2015; accepted July 18, 2015. Date of publication August 04, 2015; date of
current version October 06, 2015. The associate editor coordinating the reviewof this manuscript and approving it for publication was Dr. John McAllister.
This material is based upon work supported by the National Science Founda-
tion (NSF) under Grant No. 1307775 and U.S. Air Force Office of ScientificResearch (AFOSR) under Grant No. FA9550-10-1-0458. A part of this work
was presented at ICASSP, Florence, Italy, 2014.The authors are with the Department of Electrical Engineering and Computer
Science, SyracuseUniversity, Syracuse, NY 13244 USA (e-mail: [email protected]; [email protected]; [email protected]; [email protected]).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSP.2015.2464197
areas including battlefield surveillance [1], environment moni-
toring and target tracking [2], industrial processes [3] and health
monitoringand control [4]. In this paper,we assumethat thetask
of the WSN is to track a target emitting energy in a given region
of interest (ROI). The sensors in the ROI report their measure-
ments to a central node called the fusion center which is respon-
sible for the final inference.
Typical WSNs have limited resources (energy, bandwidth),
and the sensors are deployed in the ROI densely. Thus, instead
of simply having all the sensors transmit all the time, proper
management and programming of a subset of sensors that
should transmit their observations is very important. Different
approaches have been proposed to solve the sensor manage-
ment problem in the literature for various inference tasks. To
name a few, in [5], the sensor selection problem was formulated
as an integer programming problem, which was relaxed and
solved through convex optimization. In [6], a multi-step sensor
selection strategy by reformulating the Kalman filter was pro-
posed, which was able to address different performance metrics
and constraints on available resources. In[7], a sensor selection
scheme based on an entropy-based information measure wasproposed. The recursive posterior Cramr-Rao lower bound
(PCRLB) on the mean squared error (MSE) has been explored
as the metric to select informative sensors in [8] and [9]. To
decide which sensors are important, the innovation of each
sensor which is defined as the difference between the current
measurement and the predicted measurement has been utilized
in [10]. In [11], the nondominating sorting genetic algorithm-II
method was employed for the multi-objective optimization
based sensor selection problem. In [12], the authors aimed
to find the optimal sparse collaboration topologies subject to
a certain information or energy constraint in the context of
distributed estimation. For a more complete literature review onsensor management for target tracking, see [13] and references
therein.
In a densely deployed WSN, since only a few nodes have
significant observations, the concatenated measurement vector
can be considered to be sparse and compressible. In [13][16],
a sparse formulation is exploited to reduce the number of se-
lected sensors. In [13], the problem of periodic sensor sched-
uling was addressed by seeking the optimal sparse estimator
gain, where a one-to-one correspondence between active sen-
sors and the nonzero columns of the estimator gain was estab-
lished. In [14], the design of the sensor selection scheme was
transformed to the recovery of a sparse matrix. In [15] and [16],
1053-587X 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
7/23/2019 07177096(1)
2/12
6050 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 63, NO. 22, NOVEMBER 15, 2015
a sparsity-aware sensor selection problem was formulated by
minimizing the number of selected sensors subject to a certain
estimation quality. The concept of compressive sensing (CS)
has been discussed in [17] and [18]. The authors in [19] in-
troduced the application of CS to radar sensor networks. They
showed that the signal samples along the time domain could be
substantially compressed so that signals could be recovered by a
small number of measurements. Also, the maximum likelihood
(ML) algorithm was developed for parameter estimation and the
Cramer-Rao lower bound (CRLB) was provided to validate the
theoretical result. In [20], a multiple target localization approach
was proposed by formulating the multiple target locations as a
sparse matrix, and the target locations were recovered from the
noisy measurements through -minimization. The first attempt
to solve the sensor management problem by CS was presented
in [21], in which the sensor selection decision was considered as
a sparse signal, and the sensor selection problem was solved in
terms of recovering the sparsesignal by norm minimization.
However, the probabilistic sensor management problem using
the compressed data from the sparse observations of the sen-sors has not been considered in the above references.
In this paper, we propose a novel CS based sensor manage-
ment approach for target tracking in a WSN. Since not all the
sensors contain informative observations regarding the target at
a given time instant, the observation vector has only a few sig-
nificant elements. Thus, it is sufficient to forward only those sig-
nificant elements to the fusion center to perform target tracking
instead of forwarding all the measurements which consumes
a large amount of energy. To get a compressed version of the
observations at the fusion center, we employ a multiple access
channel (MAC) with probabilistic transmissions. Use of a MAC
model to get a compressed version of observations has been
discussed by several authors, for example, [22] and [23]. In
this approach, each sensor multiplies its measurement with a
random scalar drawn from a given distribution and transmits
it via a MAC. Then, the received observation vector at the fu-
sion center has an equivalent representation as with the stan-
dard CS problem. With this model, the corresponding sensing
matrix at the fusion center is completely determined by each
sensors probability of transmission. Under this framework, our
goal is to design the sensing matrix considering two cases: 1)
the fusion center has complete measurement sparsity informa-
tion (CMSI), and, 2) the fusion center has partial measurement
sparsity information (PMSI). With CMSI, the fusion center gen-
erates the sensing matrix based on the transmission probabilitiesof the sensors. With PMSI, the fusion center sends the trans-
mission probabilities to the sensors once it gets the optimal so-
lutions. Then the sensors generate the sensing matrix based on
their own transmission probabilities and decide whether or not
they should transmit their observations to the fusion center. It
is noted that, with CMSI, the fusion center has to inform which
sensor should transmit during each MAC transmission, while
with PMSI, the fusion center sends only the probability values
to each node which is the same for all MAC transmissions.Thus,
in terms of the communication burden required by the feed-
back channel (fusion center to sensor nodes), the PMSI based
method is more efficient than the CMSI method. However, sim-ulation results show that the CMSI method is capable of pro-
viding better tracking performance compared to the PMSI based
method when the processing noise is relatively large. Under
both schemes, we obtain the optimal values of transmission
probabilities that generate the measurement matrix so that a
given performance metric for target tracking is optimized.
An initial version of this work was reported in [24], where we
studied only the CMSI case. In the current work, we extend the
work reported in [24] in several directions:
We develop a probabilistic sensor management scheme for
target tracking based on compressed measurements con-
sidering PMSI in addition to the CMSI case as considered
in [24].
When the sensor nodes transmit their observations with
nonidentical probabilities, we formulate the sensor man-
agement problem as a quasi-convex optimization problem
and obtain the optimal values via a linear program for both
CMSI and PMSI cases.
With identical probabilitiesof transmission, the theoretical
results provide us with some intuitive insights into the im-
pact of the energy constraint, the number of sensors in theROI, and the number of MACs on the performance of the
CS based target tracking problem.
While a CS based sensor management approach for WSNs
has been discussed in [21], there are several major differences
between our work and the work presented in [21]: 1) The authors
in [21] considered a linear system while our model is nonlinear
and is, thus, more general. 2) In [21], a subset of sensors is se-
lected and the selected sensors send their measurements to the
fusion center over parallel channels. In this paper, a subset of
sensors is chosen probabilistically and different superpositions
of weighted measurements are sent to the fusion center over
MACs. 3) In [21], the sensor selection decision is considered
as a sparse signal and the sensor selection problem is solved by
recovering the sparse signal by norm minimization. In this
paper, the concatenated measurement vector is considered to be
sparse due to the presence of non-informative measurements,
and the sensing matrix is designed such that a desired tracking
performance is achieved with compressed measurements. Thus,
there is no recovery of signal, but the compressed signal is used
directly for state inference; 4) In [21], the sensing matrix is de-
terministic or is made semi-random by adding some random
disturbance, while in this paper, elements of the sensing matrix
are random variables whose distributions are related to sensors
probabilities of transmission.
The rest of the paper is organized as follows. In Section II,we introduce the formulation of our problem. In Section III, par-
ticle filtering based Fisher information matrix (FIM) for CMSI
and PMSI are derived. We present the optimization problem for
probabilistic sensor management in Section IV and the numer-
ical experiments in Section V, respectively. Our work is con-
cluded in Section VI.
II. PROBLEM FORMULATION
A. System Model
We consider a WSN consisting of sensors which are de-ployed uniformly in a square region of interest (ROI) of size
7/23/2019 07177096(1)
3/12
ZHENGet al.: COMPRESSIVE SENSING BASED PROBABILISTIC SENSOR MANAGEMENT FOR TARGET TRACKING IN WSNs 6051
. Note that our approach can handle any sensor deploy-
ment pattern as long as the sensor locations for all
are known in advance. We assume that the target and
all the sensors are based on a flat ground, so that we can formu-
late the problem with a 2-D model. We focus on a target tracking
problem, where a moving acoustic or electromagnetic target is
tracked by the WSN. The dynamics of the target is defined by a
4-dimensional state vector where
is the location of the target at time instant and , are the
velocities in the and directions. The model of the target mo-
tion is assumed to be
(1)
where is the state transition model and is the process noise
which is assumed to be Gaussianwith mean zero and covariance
matrix .
At time , the measurement model at each sensor is
(2)
where , is the signal power of the
source, is the signal decay parameter, denotes the dis-
tance between the target and the th sensor at time , i.e.,
, where is the loca-
tion of the th sensor, and is the measurement noise, which
is assumed to be Gaussian with mean zero and variance and
mutually independent over for .
Let the measurement vector be at time
, where denotes the matrix or vector transpose. We con-
sider a relatively large distributed network. Based on the obser-
vation model (2), it is seen that the signal amplitude received
at a given node at a given time becomes smaller and eventu-
ally negligible as the distance between that particular node and
the true target location increases. Therefore, at time ,
contains only a few significant values. This
motivates us to consider a scheme where only a compressed
version of is transmitted to the fusion center instead of the
complete observation vector .
B. Compressive Sensing (CS)
For a densely deployed WSN, the signal measurements are
considered to be sparse and compressible. CS is a recently de-
veloped signal processing technique for acquiring and recon-
structing a sparse signal with a small number of measurements
compared to the original signal dimension. Consider a signal, that can be expressed in an orthonormal basis
as
where is the coefficient of the signal projected on to and
. The signal is said to be -sparse, if only
coefficients in are significant and all the others are zeros or
negligible.
To get a compressed signal, the sparse signal is projected
to a lower dimension via a sensing matrix with dimension
, where , i.e.,
Then, the standard CS problem is to recover from only
measurements . The reconstruction capability is deter-
mined by the properties of the sensing matrix in addition to
the sparsity index and the number of compressed measurements
[17]. Several such properties including restricted isometry prop-
erty (RIP) and mutual coherence of the sensing matrix, and re-
covery algorithms are discussed in [25][27].
C. Sparsity Formulation of Sensor Management Problem
To obtain a compressed version of observations at the fusion
center, we consider the following transmission scheme as con-
sidered in [23]. In particular, to get a compressed version of
the observations at the fusion center, we employ a multiple ac-
cess channel (MAC) based communication scheme with prob-
abilistic transmissions. In this approach, each sensor multiplies
its measurement with a random scalar, which denotes whether
or not to transmit its measurement to the fusion center, drawn
from a given distribution and transmits it via a MAC. Then, the
received observation vector at the fusion center is a compressed
version of the original observation vector and has an equiva-lent representation as in the standard CS problem. Let the th
sensor transmit its measurement after multiplying it by (to
be defined later) via a MAC, so that after transmissions, the
received signal at the fusion center is given by
(3)
where is the receiver noise, which is assumed to be white
and Gaussian with mean zero and variance . Note that (3) can
be written in a vector form as
(4)
where , the -th element of is
given by for and , and is
the receiver noise, which is assumed to be white Gaussian with
mean zero and covariance matrix , where
is an identity matrix of size .
We consider each to be a random variable so that
with prob.
with prob.
with prob.
(5)
where is the probability of transmission of th sensor attime instant .
Based on how is constructed, it is obvious that, though the
elements in a given column in are independent and identi-
cally distributed (i.i.d.), elements in different columns are inde-
pendent but not identically distributed. It is noted that, the ma-
trix can be very sparse when only a small number of sensors
decide to transmit with a high probability. With this sensing ma-
trix, we show numerically in Section V that compressed obser-
vations in (4) provide us with tracking performance comparable
to that with (2) with relatively small .
In the context of sensor management, plays the role of
a sensor management entity that divides the sensors into
subsets. Sensors in the same subset send their superimposed
measurements over the same MAC (there are a total of
7/23/2019 07177096(1)
4/12
6052 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 63, NO. 22, NOVEMBER 15, 2015
MACs) if FDMA is used or in the same time slot (there are a
total of time slots) if TDMA is used. Note that, the weight
could be 0, which means that the associated sensor does not
send its measurement. In this scenario, the sensor observations
are compressed from dimension to , and the compression
is achieved as a result of coherent transmission over MAC chan-
nels (either with FDMA or TDMA). Therefore, the problem of
managing sensors is equivalent to the design of the sensing ma-
trix or the probability vector ,
such that a certain objective function is optimized. In our paper,
we choose the Fisher information, which is the inverse of
the bound on the mean squared error of the estimator namely
PCRLB, as our optimization criterion, i.e., we address this
probabilistic sensor management problem so that the trace of
the FIM is maximized at each time step .
III. FISHERINFORMATION MATRIX FORCS BASED
TARGET TRACKING
In our framework, the fusion center decides the transmission
probabilities of the sensors, and then communicates with the
sensors to get the compressed data through the MACs. Thus,
regarding the communication scheme between the fusion center
and the sensors, the natural question is, who decides the sensing
matrix in (4)? We consider two different cases: 1) the fusion
center generates the sensing matrix and informs the sensors as
to which sensors are to transmit during each MAC transmission,
in which case the fusion center has CMSI; and, 2) the fusion
center sends only the transmission probabilities of the sensors,
in this case, the sensors generate the sensing matrix and decide
how they send their observations to the fusion center, i.e., the
fusion center has PMSI. Therefore, before proceeding with the
optimization problem, in this section, we find the FIM as a func-tion of the transmission probabilities of the sensors for the two
different transmission schemes.
For the target tracking problem under consideration, a nice
recursive computation of the FIM at time , , is pro-
posed in [28], which is given as follows
(6)
where is the FIM at time , and
In the above expressions for , , , , denotes
the probability density function, is the state of the target
at time , is the measurement vector at time ,
denotesthe second derivative operator, namely ,
and isthe gradient operator, and denotes the mathemat-
ical expectation.
For the problem considered in this paper, we have
, , . To compute
, it is required to compute . It is noted that
depends on the random matrix . Here we consider two
different cases: 1) the fusion center generates the sensing matrix
and decides the selection state of each sensor,so that it has
complete information about the sensing matrix . Then the
conditional probability can be applied to
obtain ; 2) the fusion center sends the transmission prob-
ability to sensor , and sensor generates
corresponding to its selection prob-
ability. In this case, the fusion center does not have the exact
values of . Thus, is computed by averaging
over the sensing matrix .
A. Fisher Information Matrix With Complete Measurement
Sparsity Information (CMSI)
With complete sparsity information, is calculated from
the conditional probability ,
(7)
where the expectation is with respect to the joint distribution
. Hence,
(8)
where
(9)
Notice that the second expectation in (9) is with respect to
. Recall (4), where we have
(10)
where , i.e., it is the concatenation
of themeasurement noises of sensors. Since themeasurement
noises are mutually independent, , where
represents the multivariate Gaussian distribution withmean vector and the covariance matrix .
Given and , based on (10), one can get
, where . Then
Therefore, in (9) reduces to
(11)
where
and
7/23/2019 07177096(1)
5/12
ZHENGet al.: COMPRESSIVE SENSING BASED PROBABILISTIC SENSOR MANAGEMENT FOR TARGET TRACKING IN WSNs 6053
for . Note that is the location of the th
sensor and represents the location of the target at
time .
Letting , we observe that it is
difficult to obtain directly. Thus, in the following propo-
sition, we employ some approximations when is large.
Proposition 1: If the number of sensors in the WSN is
large, then, at any given time , we may approximate
where denotes a diagonal matrix, with on the
main diagonal.
Proof: See Appendix A.
We see that Proposition 1 yields a diagonal formulation of
, which makes the relationship between and
much more clear. Thus, we have
(12)
(13)
where
Note that in (13) we take the expectation of because
is the only element that is related to the target state
. Therefore, the FIM for the CMSI case is computed by
substituting in (13) back to (6).
B. Fisher Information Matrix With Partial Measurement
Sparsity Information (PMSI)
In this subsection, we consider the second case where the sen-
sors generate sensing matrix based on by them-
selves and the fusion center does not have the exact information
regarding . In this case, to compute ,
is computed as
With similar approximations as considered in Section III-A, we
have the following proposition.
Proposition 2: If the number of sensors is large, then, at
any given time , we may approximate
(14)
Proof: See Appendix B.
Hence,
(15)
(16)
where is a 4 4 matrix with
and
All the other elements in are zeros. Also note that in (16) we
take the expectation of because is the only
element thatis related tothetarget state . Thus, the FIM for
the PMSI case is computed by substituting in (16) back
to (6).
C. Particle Filter Based Fisher Information Matrix
After computing the FIM for CMSI and PMSI cases, in this
subsection, we briefly introduce particle filters and give simpler
expressions of the FIM with the help of particle filters. Since
particle filtering is a general filtering procedure that can be ap-
plied to any state-space model and thus is more general than
the Kalman filter [29], we apply particle filtering for our non-
linear model to track the target. In particle filtering, the main
idea is to find a discrete representation of the posterior distri-
bution by using a set of particles with as-
sociated weights . The posterior density at can be
approximated as,
7/23/2019 07177096(1)
6/12
6054 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 63, NO. 22, NOVEMBER 15, 2015
where is the Dirac delta measure and denotes the total
number of particles. We employ the sequential importance re-
sampling (SIR) particle filtering algorithm [29] to solve the non-
linear Bayesian filtering problem. Algorithm 1 provides a sum-
mary of SIR particle filtering for sensor selection. A more de-
tailed treatment of particle filtering can be found in a wide va-
riety of publications such as [29].
Algorithm 1:SIR Particle Filter with Probabilistic Sensor
Management for Target Tracking
Set . Generate initial particles with.
while do
a) Propagating particles:
b) Obtain sensor data based on the sensormeasurements through the MAC of transmissions.
c) Updating weights:
(for CMSI)
(for PMSI)
d) Normalizing weights:
e) Target Estimation:
f) Resample:
end while
In Algorithm 1, denotes the number of time steps over
which the target is tracked. For the first case, i.e., CMSI,
is according to (10),
and for the second case, i.e., PMSI, is approxi-
mated by (14). The resampling step avoids the situation that all
but one of the importance weights are close to zero [29].
With particle filtering, expectation of in (13) and
expectation of in (16) can be written as
(17)
and
(18)
Thus, we provide the FIM of our system in the following
proposition.
Proposition 3: The FIM of our system for both the CMSI and
PMSI can be written as
(19)
where is different for the two cases,
for CMSI
for PMSI.
Proof: We get (19) by substituting (17) and (18) back into
(13) and (16).
IV. THESENSORMANAGEMENT PROBLEM
Having computed the FIM in Section III for both CMSI and
PMSI cases, in this section, we focus on our CS based sensor
management problem in the WSN. We observe that for our
model, the problem of maximizing the determinant of the FIMin
(19) subject to the resource constraint is not convex, thus global
optimal solutions are not guaranteed, and the computation costs
will be high. Therefore, we solve the CS based resource man-
agement problem in a WSN by maximizing the trace of theFIM.
For simplicity, we assume that each transmission from a local
sensor to the fusion center consumes unit power. Finding the op-
timal values for transmit power at sensor nodes while achievinga desired performance is an interesting aspect which will be
studied in the future. We aim to solve the following optimiza-
tion problem:
(20a)
(20b)
(20c)
where denotes the trace of a matrix, is the total energy
constraint, and is the lower bound of the selection probability.
Here we impose a lower bound on the probability to avoid pro-ducing columns with all zeros in the measurement matrix .
In other words, we allow each node to transmit with a certain
nonzero probability.
At time step , the fusion center first solves the optimiza-
tion problem in (20) to obtain the optimal before measure-
ments at this time are available. Then, 1) For the CMSI case
in III-A, the fusion center generates the sensing matrix
using , and, according to which, sends control messages
to local sensors. Based on these control messages, local sen-
sors send their measurements over assigned MACs to the fu-
sion center. For the PMSI case in III-B, the fusion center sends
the transmission probability to sensor .Based on its selection probability , sensor generates
, according to which it sends its mea-
surements to the fusion center.
A. With Identical Probabilities
We first consider a simple scenario in which each sensor has
the same probability to be selected to transmit their measure-
ments. Since from (19) we know that does not depend on
the transmission probability of the sensors, the FIM for both the
CMSI and PMSI cases with the assumption of identical trans-
mission probability is
7/23/2019 07177096(1)
7/12
ZHENGet al.: COMPRESSIVE SENSING BASED PROBABILISTIC SENSOR MANAGEMENT FOR TARGET TRACKING IN WSNs 6055
The objective function of the optimization problem (20a) is
Thus, no matter whether or not the fusion center has completeinformation about the sensing matrix , the resource manage-
ment problem of (20) is equivalent to
(21)
Since is an increasing function of , the
optimal solution is simply the largest value that can achieve
under the constraints
(22)
We observe that the optimal solution for such a simple sce-
nario is no longer dependent on the FIM of each sensor, i.e.,
this formulation treats all the sensors to be equally important,
even though they make different contributions for tracking the
target. When we do not consider the case that every sensor has
probability 1 to transmit its measurements, we get some insights
from (22): 1) A larger energy constraint results in a higher
transmission probability for the sensors. 2) In the identical prob-
ability case, having more sensors in the WSN yields a lower
transmission probability for the sensors. It is because all the sen-
sors are forced to transmit their measurements with the sameprobability. 3) In the context of CS, is the number of com-
pressed measurements. A larger is supposed to yield a higher
probability to recover the original signal, so that better perfor-
mance is expected. However, on the other hand, in (20b), the
value of also decides the upper bound of the total selection
probability of the sensors, thus increasing decreases the sen-
sors transmission probability as shown in (22), which may yield
worse tracking performance. Therefore, there exists a tradeoff
between the accuracy of the recovered signal and tracking per-
formance.
B. With Nonidentical Probabilities
Without the simplifying assumption that each sensor employs
an equal transmission probability, our optimization problem is,
(23)
Note that
for CMSI
for PMSI.
By denoting , the objective func-
tion is further expressed as
(24)
where is a unit vector, . The
objective function in (24) is a linear-fractional function, so that
the optimization problem is a quasilinear optimization problem
[30]. It has been shown in [30] that, as the feasible set
is not nonempty, the linear-fractional problem (24) can be trans-
formed to an equivalent linear program. We first define the fol-
lowing two variables
such that once we get the optimal solutions for and , we can
also easily get the optimal values for the original variables .
The equivalent linear program optimization problem is
(25)
where
......
. . ....
......
. . ....
We apply the MATLAB function linprog to solve the con-strained linear optimization problem (25). Having the optimal
solutions and , the transmission probabilities of the sensors
are .
V. SIMULATION RESULTS
In this section, we illustrate the performance of the proposed
sensor management algorithm by numerical examples. We com-
pare the MSE of CMSI and PMSI presented in Section III-A and
Section III-B, respectively, with different process noise param-
eters. The MSE is computed through
7/23/2019 07177096(1)
8/12
6056 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 63, NO. 22, NOVEMBER 15, 2015
Fig. 1. Target trajectory and the estimated track with different sensor manage-
ment schemes.
where is the number of Monte Carlo trials. Also, the effect
of the parameters in the model, i.e., the energy constraint ,
the number of sensors and the number of MACs ( of the
sensing matrix ) is evaluated.
We consider a WSN, consisting of sensors grid deployed
in a surveillance area. The dynamical
model of the target is given by (1). The state transition model
and the covariance of the process noise are given as follows:
where is the sampling time interval and isthe process noise parameter. The parameters of the observation
model (2) are set as and . The initial state
of the target is assumed to be Gaussian with mean
and covariance matrix .
We perform target tracking over time steps for each
Monte-Carlo trial, and set particles for the particle
filter. The MSE of the estimation at each time is averaged over
Monte-Carlo trials.
In Fig. 1, a WSN is illustrated where sensors track
a target with MACs under the energy constraint
. We plot the true target trajectory, the estimated target
trajectory when all the sensors send their measurements to thefusion center, the estimated target trajectory with the proposed
CS based sensor management method when the FC has CMSI
and the transmission probabilities of the sensors are noniden-
tical, and the estimated target trajectory under the random selec-
tion scheme. For the random selection scheme, the sensors are
randomly selected by the fusion center with probability .
Thus, the random selection method is expected to give worse
tracking performance compared to the CS based sensor man-
agement method and the all-send case. It is also observed that
the proposed CMSI based scheme shows similar performance
as that of the all-send case.
In Fig. 2, we study the tracking performance with CMSI
and PMSI respectively. With sensors in the ROI, we
compare the MSE under different process noise parameters.
Fig. 2. Comparing CMSI and PMSI via MSE with different noise parameters.
Note that under the identical probability condition, sensors
are equally important, so that whether the fusion center has
CMSI or PMSI does not affect the tracking performance. We
observe in Fig. 2(a) that when the process noise is small
, PMSI outperforms CMSI a little bit, and in
Fig. 2(b) when the noise is relatively large , CMSI
performs much better than PMSI. The reason is that when the
process noise is small, the target motion model has relatively
smaller uncertainty about the predicted target location. Thus,
averaging the sensing matrix helps us get even better tracking
performance. However, when the process noise is relatively
large, the fusion center has higher uncertainty about the target
trajectory. Moreover, without having complete information
about the sensing matrix, the fusion center calculates the
weights of the particles in particle filtering with errors. There-fore, the fusion center is able to achieve better performance
with CMSI compared to PMSI.
Now we study the effect of the parameters in the WSN (the
energy constraint , the number of sensors and the numberof
MACs) on the CS based sensor management problem. The per-
formance is evaluated through the tracking performance under
the assumption that the process noise is , and that the
fusion center has CMSI to track the target.
In Fig. 3, the tracking performance is illustrated when
sensors track a target with MACs under
different energy constraints . In Fig. 3, we give the MSE
for and , to show the efficiency of our model,where Fig. 3(a) shows the MSEs when the sensors have non-
identical probabilities, and Fig. 3(b) shows the MSEs when
the sensors have identical probabilities. We also present the
MSE for the random selection method for both cases. Note
that for the random selection scheme, the sensors are randomly
selected with probability in Fig. 3(a) and with probability
in Fig. 3(b) corresponding to the optimal solution in
(22). From both figures, we observe that when increases from
6 to 8, the MSE decreases for both the random selection and
the CS based approach, and the latter one always outperforms
the former under the same energy constraint. Both methods are
compared to the all-send case where all sensor measurements
are available at the fusion center via a set of parallel channels.
Compared to the all-send case where a total of 25 units of
7/23/2019 07177096(1)
9/12
ZHENGet al.: COMPRESSIVE SENSING BASED PROBABILISTIC SENSOR MANAGEMENT FOR TARGET TRACKING IN WSNs 6057
Fig. 3. MSE for different energy constraints, , 8 (a) Nonidentical probabilities. (b) Identical probabilities.
Fig. 4. MSE for different number of sensors, , 25, 36, 49 (a) Nonidentical probabilities. (b) Identical probability.
energy are consumed at each time since parallel
transmissions are necessary, the proposed approach loses only
a little performance for both and cases.
In Fig. 4, we show the tracking performance for different
number ofsensors , 25,36, 49withthe energy constraint
and the number of MACs , where Fig. 4(a) shows
the MSEs when the sensors have nonidentical probabilities, and
Fig. 4(b) shows the MSEs when the sensors have identical prob-
abilities.In Fig. 4(a),the MSEs show that theproposed approach
achieves a better performance as increases. This is because,in target tracking without the strict constraint that every sensor
is equally important, the sensors which can obtain more infor-
mative observations in the ROI will be assigned higher proba-
bilities. On the other hand, as the number of sensors in the ROI
increases, the density of the sensors increases, so that the fu-
sion center has a better chance to activate even more informative
sensors. When the density is large enough, the tracking perfor-
mance tends to saturate as shown in Fig. 4(a). However, from
Section IV-A, we know that under the assumption that all the
sensors have identical transmission probabilities, the transmis-
sion probability of the sensors becomes smaller as the number
of sensors in the ROI increases, so that the tracking performance
worsens as shown in Fig. 4(b). Note that in Fig. 4, in order to
clearly see that the number of sensors in the WSN affects the
tracking performance, we have set the number of MACs to
be relative small, which results in divergence of the MSE espe-
cially when is relatively small. We shows the effect of the
number of MACs in Fig. 5.
In Fig. 5, the MSEs of the CS based approach with different
values are presented, where Fig. 5(a) shows the MSEs when
the sensors have nonidentical probabilities, and Fig. 5(b) shows
the MSEs when the sensors have identical probabilities. In
Fig. 5, we let and . As observed before, it is
seen from Fig. 5(a) and Fig. 5(b), that the tracking performanceis improved when nonidentical probabilities are assigned to
sensors compared to assigning identical probabilities under
the same energy constraints and the same number of MAC
transmissions. In both cases (identical and nonidentical proba-
bilities) the performance is improved as M increases. However,
it is observed that the rate of performance improvement reduces
as M increases. It is noted that, in Section IV-A, we indicated
that with the identical transmission probability assumption, the
transmission probability of the sensors decreases as M increases
due to the energy constraint. However, the following reasons
justify the improvement of the tracking performance as M
increases (even though the transmission probability decreases):
1) As M, the number of MAC transmissions, increases the
fusion center receives different superpositions of the observa-
7/23/2019 07177096(1)
10/12
6058 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 63, NO. 22, NOVEMBER 15, 2015
Fig. 5. MSEs for different number of MACs, , 3, 4 (a) Nonidentical probabilities. (b) Identical probability.
tions from the sensors with good observations. This enables
the fusion center to better extract the most informative data
residing at distributed nodes. 2). M is the dimension of the
compressed observation vector received at the fusion center
in (4). In the context of CS with respect to complete signal
recovery, it has been shown in [31] that reliable recovery is
possible with a small number of compressive measurements
(compared to original signal dimension) even if the projection
matrix contains sparse random elements as long as the matrix
satisfies some required properties. Then, increasing M beyond
this particular value is not necessary. While these analyses
(e.g., [31]) may not be directly applicable to the scenario
considered in this paper, (where we consider tracking instead of
complete recovery when the random projection matrix contains
sparse nonidentical elements), it is intuitive that when M isincreased beyond a certain value, a significant improvement in
performance is not observed. The performance of the proposed
approach is quite close to the all send case, especially when
, 4, but is energy efficient in the sense that it consumes
only units of energy at any given time on an average.
Note that the compressed measurements are used directly for
state estimation in the considered target tracking problem,
which is different from the traditional CS problems where the
goal is to recover a sparse signal.
VI. CONCLUSION
In this paper, we proposed a novel probabilistic sensor man-
agement approach for target tracking in sensor networks based
on compressed observations. With the proposed approach, the
sensor management problem becomes a constrained optimiza-
tion problem, where the goal is to determine the optimal values
of probabilities that each sensor should transmit with such that
the trace of the FIM at any given time step is maximized. We de-
rived the formulations for the FIM when the fusion center has
complete measurement sparsity information (CMSI) and partial
measurement sparsity information (PMSI). Theoretical results
for the case when each sensor employs identical transmission
probability give us useful insights into the impact of the three
important parameters of our model. Numerical results for both
the identical probability case and the more general nonidentical
transmission probabilities case show that the proposed approach
saves a lot of energy with a little performance loss compared to
the optimal scenario in which all sensor observations are trans-
mitted to the fusion center via parallel channels. Under the same
energy constraint, the proposed scheme outperforms the random
selection approach significantly. An interesting future work is to
take the channel statistics into consideration.
APPENDIX A
PROOF OF PROPOSITION 1
Let . Note that in this proof the time index is
omitted for the sake of simplicity. Diagonal elements of are
given by
and off-diagonal elements are
It is straightforward to obtain ,
, , and
.
Therefore, according to the law of large number (LLN) for
independent and nonidentical random variables, we get
Hence, and
(26)
Then,
7/23/2019 07177096(1)
11/12
ZHENGet al.: COMPRESSIVE SENSING BASED PROBABILISTIC SENSOR MANAGEMENT FOR TARGET TRACKING IN WSNs 6059
Diagonal elements of are given by
and off-diagonal elements are
Therefore,
completing the proof.
APPENDIX B
PROOF OF
PROPOSITION
2Based on Eq. (10), one can get
, where .
Then we have
With LLN, we have the approximation for as in (26).
Then
and
(27)
According to LLN, we can get approximations
(28)
and
(29)
Therefore, the proof is completed by substituting (28) and (29)
into (27).
REFERENCES
[1] T. Bokareva, W. Hu, S. Kanhere, B. Ristic, N. Gordon, and T. Bessell
et al., Wireless sensor networks for battlefield surveillance, in Proc.
Conf. Land Warfare, 2006, pp. 18.
[2] J. Yick, B. Mukherjee, and D. Ghosal, Wireless sensor network
survey,Comput. Netw., vol. 52, no. 12, pp. 22922330, 2008.
[3] V. C. Gungor and G. P. Hancke, Industrial wireless sensor networks:
Challenges, design principles, technical approaches,IEEE Trans. Ind.Electron., vol. 56, no. 10, pp. 42584265, 2009.
[4] A. Milenkovi, C. Otto, and E. Jovanov, Wireless sensor networks for
personal health monitoring: Issues and an implementation, Comput.
Commun., vol. 29, no. 13, pp. 25212533, 2006.
[5] S. Joshi and S. Boyd, Sensorselectionvia convex optimization,IEEE
Trans. Signal Process., vol. 57, pp. 451462, Feb. 2009.
[6] Y. Mo, R. Ambrosino, and B. Sinopoli, Sensor selection strategies
for state estimation in energy constrained wireless sensor networks,
Automatica, vol. 47, pp. 13301338, Jul. 2011.
[7] F. Zhao, J. Shin, and J. Reich, Information-driven dynamic sensor
collaboration, IEEE Trans. Signal Process., vol. 19, no. 1, pp. 6172,
May 2002.
[8] L. Zuo, R. Niu, and P. Varshney, Posterior crlb based sensor selec-
tion for target tracking in sensor networks, in Proc. IEEE Int. Conf.
Acoust., Speech, Signal Process., Apr. 2007, pp. 10411044.
[9] L. Zuo, R. Niu, and P. Varshney, A sensor selection approach for
target tracking in sensor networks with quantized measurements, in
Proc. IEEE I nt. Conf. Acoust., Sp eech, Signal Process., Mar. 2008, pp.
25212524.
[10] G. Wang, J. Chen, J. Sun, and Y. Cai, Power scheduling of Kalman
filtering in wireless sensor networks with data packet drops, 2013,
arXiv preprint arXiv:1312.3269.
[11] N. Cao, E. Masazade, and P. K. Varshney, A multiobjective optimiza-
tion based sensor selection method for targettracking in wireless sensor
networks, inProc. IEEE Int. Conf. Inf. Fusion (FUSION), 2013, pp.
974980.
[12] S. Liu, M. Fardad, S. Kar, and P. Varshney, On optimal sensor collab-
oration topologies for linear coherent estimation, in Proc. IEEE Int.
Symp. Inf. Theory (ISIT), 2014, pp. 26242628.
[13] S. Liu, M. Fardad, E. Masazade, and P. Varshney, Optimal periodic
sensor scheduling in networks of dynamical systems, IEEE Trans.Signal Process., vol. 62, no. 12, pp. 30553068, Jun. 2014.
[14] I. D. Schizas, Distributed informative-sensor identification via spar-
sity-aware matrix decomposition,IEEE Trans. Signal Process., vol.
61, no. 18, pp. 46104624, 2013.
[15] S. Chepuri and G. Leus, Sparsity-promoting sensor selection for non-
linear measurement models,IEEE Trans. Signal Process., vol. 63, no.
3, pp. 684698, Feb. 2015.
[16] S. Chepuri and G. Leus, Sparsity-promoting adaptive sensor selec-
tion for non-linear filtering, in Proc. IEEE Int. Conf. Acoust., Speech,
Signal Process. (ICASSP), May 2014, pp. 50805084.
[17] D. Donoho, Compressed sensing, IEEE Signal Process. Mag., pp.
118124, Jul. 2007.
[18] E. Cands and M. B. Wakin, An introduction to compressive sam-
pling, IEEE Signal Process. Mag., vol. 25, no. 2, pp. 2130, Mar.
2008.
[19] L. Xu, Q. Liang, X. Cheng, and D. Chen, Compressive sensing in dis-
tributed radar sensor networks using pulse compression waveforms,
EURASIP J. Wireless Commun. Netw., vol. 2013, no.1, pp.110, 2013.
[20] C. Feng, S. Valaee, and Z. Tan, Multiple target localization using
compressive sensing, in Proc. IEEE Global Telecommun. Conf.
(GLOBECOM), 2009, pp. 16, IEEE.
[21] A. Carmi and P. Gurfil, Sensor selection via compressed sensing,
Automatica, pp. 33043314, Nov. 2013.
[22] J. Haupt, W. U. Bajwa, M. Rabbat, and R. Nowak, Compressed
sensing for networked data,IEEE Signal Process. Mag., vol. 25, no.
2, pp. 92101, 2008.
[23] G. Yang, V. Y. Tan, C. K. Ho, S. H. Ting, and Y. L. Guan, Wireless
compressive sensing for energy harvesting sensor nodes, IEEE Trans.
Signal Process., vol. 61, no. 18, pp. 44914505, 2013.
[24] Y. Zheng, T. Wimalajeewa, and P. K. Varshney, Probabilistic sensor
management for target tracking via compressive sensing, in Proc.IEEE Int. Conf. Acoust., Speech , Signal Process. (ICASSP), May 2014,
pp. 50955099.
7/23/2019 07177096(1)
12/12
6060 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 63, NO. 22, NOVEMBER 15, 2015
[25] E. Cands, J. Romberg, and T. Tao, Robust uncertainty principles:
Exact signal reconstruction from highly incomplete frequency infor-
mation, IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489509, Feb.
2006.
[26] E. C. T. Tao, Near-optimal signal recovery from random projections:
Universal encoding strategies?,IEEE Trans. Inf. Theory, vol. 52, no.
12, pp. 54065425, Dec. 2006.
[27] J. Tropp and A. Gilbert, Signal recovery from random measurements
via orthogonal matching pursuit,IEEE Trans. Inf. Theory, vol. 53, no.12, pp. 46554666, Dec. 2007.
[28] P. Tichavsky, C. Muravchik, and A. Nehorai, Posterior Cram-Rao
bounds for discrete-time nonlinear filtering, IEEE Trans. Signal
Process., vol. 46, no. 5, pp. 13861396, May 1998.
[29] M. S. Arulampalam, S. Maskell, N. Gordon, and C. Tim, A tutorial on
particle filters for online nonlinear/non-Gauss ian Bayesian tracking,
IEEE Trans. Signal P rocess., vol. 50, no. 2, pp. 174188, Feb. 2002.
[30] S. Boyd, Convex Optimization. Cambridge, U.K.: Cambridge Univ.
Press, 2004.
[31] D. Achlioptas, Database-friendly random projections: Johnson-Lin-
denstrauss with binary coins,J. Comput. Syst. Sci., vol. 66, no. 4, pp.
671687, 2003.
Yujiao Zheng (S13) received her B.S. degree
in electronic engineering and information sciencefrom University of Science and Technology of
China (USTC) in 2008, M.S. degree in electricalengineering and Ph.D. degree in electrical and
computer engineering from Syracuse University,
Syracuse, NY, in 2011 and 2014 respectively. Herresearch interests are in the areas of statistical signal
processing with its application in target tracking,sensor management, and compressive sensing.
She received the Best Student Paper Award at theThirteenth International Conference on Information Fusion in 2010.
Nianxia Cao (S12) received the B.S. and M.S.degrees in control theory and control engineering
from Northwestern Polytechnical University (NPU),Xian, China, in 2009 and 2012. She has been
pursuing the Ph.D. deg ree in the Department of Elec-trical Engineering and Computer Science, Syracuse
University since 2012. Her research interests include
statistical signal processing, sensor managementin the wireless sensor networks, game theory, and
mechanism design.
Thakshila Wimalajeewa(M15) received the B.Sc.degree in electronic and telecommunication engi-
neering with First Class Honors from the University
of Moratuwa, Sri Lanka, in 2004 and the M.S., andPh.D. degrees in electrical and computer engineering
from the University of New Mexico, Albuquerque,NM in 20 07 and 2009, respectively.
She spent 20102012 as a postdoctoral researchassociate in theDepartmentof Electrical Engineering
and Computer Science, Syracuse University (SU),Syracuse, NY. She currently holds a research faculty
position at SU. Her research interests lie in the broad areas of communication
theory, signal processing and information theory. Her current research focuseson compressive sensing, low dimensional signal processing for communication
systems, blind signal recognition, and resource optimization in distributedsensor networks.
Pramod K. Varshney (S72M77SM82F97)
was born in Allahabad, India, on July 1, 1952. He
received the B.S. degree in electrical engineeringand computer science (with highest honors), and
the M.S. and Ph.D. degrees in electrical engineeringfrom the University of Illinois at Urbana-Champaign
in 1972, 1974, and 1976 respectively.From 1972 to 1976, he held teaching and research
assistantships with the University of Illinois. Since1976, he has been with Syracuse University, Syra-
cuse, NY, where he is currently a Distinguished
Professor of Electrical Engineering and Computer Science and the Directorof CASE: Center for Advanced Systems and Engineering. He served as the
Associate Chair of the department from 1993 to 1996. He is also an AdjunctProfessor of Radiology at Upstate Medical University, Syracuse. His current
research interests are in distributed sensor networks and data fusion, detectionand estimation theory, wireless communications, image processing, radar
signal processing, and remote sensing. He has published extensively. He is the
author of Distributed Detection and Data Fusion (New York: Springer-Verlag,1997). He has served as a consultant to several major companies.
Dr. Varshney was a James Scholar, a Bronze Tablet Senior, and a Fellowwhile at the University of Illinois. He is a member of Tau Beta Pi and is the
recipient of the 1981 ASEE Dow Outstanding Young Faculty Award. He waselected to the grade of Fellow of the IEEE in 1997 for his contributions in the
area ofdistributed detectionand data fusion. Hewas theGuest Editor of the Spe-
cial Issue on Data Fusion of the IEEE PROCEEDINGS January 1997. In 2000, he
received the Third Millennium Medal from the IEEE and Chancellors Citation
for exceptional academic achievement at Syracuse University. He is the recip-ient of the IEEE 2012 Judith A. Resnik Award. He is on the Editorial Boards of
theJournal on Advances in Information Fusionand IEEE SIGNALPROCESSINGMAGAZINE. He was the President of International Society of Information Fusion
during 2001.