I.

INTRODUCTION

Recent improvements in battery, microcontroller,

Adaptive Mobile Sensor and sensor technologies have resulted in the
development of autonomous vehicles that are
Positioning for Multi-Static inexpensive, dependable, and simple to operate.
Unmanned air vehicles (UAVs) have well-known
Target Tracking defense applications and are useful in many civilian
applications as well, such as border and coast patrol,
fire perimeter monitoring, search and rescue, etc.
UAVs are particularly well suited for situations that
are too dangerous for direct human monitoring. While
PENGCHENG ZHAN a single autonomous UAV can provide information
DAVID W. CASBEER
that is otherwise unattainable, cooperation among
Brigham Young University a team of UAVs can dramatically improve sensing
A. LEE SWINDLEHURST, Fellow, IEEE
performance [15]. The mobility of each UAV within
University of California at Irvine the team can be exploited to adaptively reconfigure
the sensing array in response to the environment and
the actions of the object being sensed [68]. It is this
additional degree of freedom we wish to exploit in a
Unmanned air vehicles (UAVs) are playing an increasingly multistatic radar tracking scenario.
prominent role in both military and civilian applications. We Using mobile and controllable sensing platforms
focus here on the use of multiple UAV agents in a target tracking
is a relatively new area of research, but has attracted
considerable attention in the generic sensor network
application where performance is improved by exploiting
literature. This type of problem is sometimes referred
each agents maneuverability. Local time-delay and Doppler
to as sensor management. The majority of work in
measurements made at each UAV are used as inputs to an
sensor management for target tracking assumes one of
extended Kalman filter (EKF) which tracks the targets position four practical observation models: range only [9, 10],
and velocity. Two simple metrics are defined to quantify the bearing only [11, 12], range/bearing [2, 13], and
accuracy of the tracking algorithm, and heading feedback to more recently, range/range-rate/bearing [14]. These
the UAVs is used to minimize the metric and improve tracking observation models are common due to the prevalent
performance. A simplified version of one of the algorithms that use of directional laser and radar sensors. However,
reduces computational complexity is also presented. Simulations small UAVs may have a difficult time making accurate
demonstrate the significant improvement that results when bearing measurements given their limited aperture
the UAV sensors are allowed to be optimally positioned during and the fact that they are much more sensitive to
tracking.
wind-blown disturbances. Moreover, power constraints
on small UAVs limit the range and strength of their
radar transmissions.
Other papers have focused on static sensors
that are capable of toggling between active and
resting states [1519] or arranging the scheduling
and reporting of measurments in an optimal manner
[20, 21]. The placement of static sensor nodes for
tracking in an underwater scenario is investigated in
Manuscript received June 20, 2007; revised March 31, 2008;
released for publication September 2, 2008.
[3], [22], [23]. The advantages of mobile sensors are
studied in [6], [24], where sensors are constrained to
IEEE Log No. T-AES/46/1/935931. remain in small groups that retain observability of the
Refereeing of this contribution was handled by W. Koch. target states. The motion and geometric formation
This work was supported by the U.S. National Science Foundation
effects of these groups on tracking performance are
under Information Technology Research Grant CCF-0428004. then investigated. In [25], the optimal positions of
a set of UAVs in a tracking scenario are determined
Authors current addresses: P. Zhan, ArrayComm LLC, San Jose,
CA; D. Casbeer, Dept. of Electrical and Computer Engineering,
based on the Cramer-Rao bound (CRB). These results
Brigham Young University, 459 Clyde Building, Provo, UT 84602, provide insight into the relationship between the
E-mail: (casbeer@byu.edu); A. L. Swindlehurst, Dept. of Electrical angular distribution of the UAV team and target
Engineering and Computer Science, University of California at state estimation using range, range-rate, and bearing
Irvine, Irvine, CA. measurements. However, the results are derived
under the assumption that the accuracy of the UAV
c 2010 IEEE
0018-9251/10/$26.00 measurements are constant, and not a function

120 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
of the relative position of the UAVs and the target.
In addition the solution proposed in [25] provides an
optimal set of UAV positions, but no path planning for
the UAVs or any guarantees that it is possible for the
dynamically constrained UAVs to reach the resulting
positions in the allotted time.
In this work, we assume a multistatic radar
scenario in which a transmitter, either airborne or on
the ground, illuminates a target with a radar signal.
A team of UAVs receives the target reflection, and Fig. 1. Multiple UAVs tracking single target.
each UAV makes a local time-delay and Doppler
measurement. These measurements are then sent
to a centralized processor or base which uses this The UAVs are assumed to be capable of velocity
collected data set to track the targets position and and altitude hold with an autopilot such as the one
velocity. Based on the calculated accuracy of the described in [26]. Therefore, for simplicity, we
tracking algorithm, the base determines trajectories consider only a two-dimensional scenario. We define
for the UAV team in order to optimize performance the position of the base station as the origin for our
and sends heading commands to the UAVs to (x, y) coordinate system where x and y correspond to
best position them for the next measurement. The the cardinal directions East and North, respectively.
maneuverability of the UAVs provides the base with Let x, y, Vx , and Vy represent, respectively, the position
the ability to improve tracking performance through of the target and the components of the target velocity
the tunable parameters in the tracking algorithm. vector in the (x, y) directions. Similarly, let xi and yi
We present two general optimization algorithms denote the position of the ith UAV, and assume that
for the problem, and a closed-form solution that each UAV is flying at the same constant speed V, with
is considerably simpler to implement but performs heading angle i measured counterclockwise from
essentially as well as the optimal approaches in many the x axis. According to [27] and after appropriate
situations. scaling, the time-delay and Doppler measurements can
The format of the paper is as follows: In be written as
q
Section II we describe the problem setup, the 1 p 2
assumed dynamic models, and relevant aspects i = x + y 2 + (x xi )2 + (y yi )2 (1)
c
of the extended Kalman filter (EKF). The issue !
of range-dependent measurement accuracy is also 1 Vx x + Vy y Vp
addressed. In Section IIIA and IIIB, two algorithms fi = p +p (2)
x2 + y 2 (x xi )2 + (y yi )2
for determining the optimal UAV heading angles
based on the EKF updates are presented. A more where
computationally efficient approach is described in
Section IIIC, based on heuristic observations of one of Vp = (Vx V cos i )(x xi ) + (Vy V sin i )(y yi ):
the algorithms in simple scenarios. Simulation results In the above equation, c and stand for the speed of
are presented in Section IV, and some conclusions are light and the wavelength of the signal, respectively.
drawn in Section V. The problem we address in this paper is how
the mobility of the UAV receivers can be exploited
II. BACKGROUND AND SYSTEM DEFINITION to improve the systems tracking performance. In
our solution to this problem, the time-delay and
A. Problem Statement Doppler estimates collected by the UAV team are
uploaded to the base and used as measurement
As illustrated in Fig. 1, we consider a multistatic updates in an EKF tracking the target. Simultaneously
radar problem in which a set of N UAVs receives the optimal heading commands for each UAV are
signal reflected by an airborne target illuminated by a calculated at the base station and then transmitted to
base transmitter. We assume that the UAVs are able to the agents. The precise optimality criteria used for
form time delay (i ) and Doppler (fi ) estimates based computing the heading commands is presented in
on these signals,1 and transmit these measurements Section III.
back to the base.
B. UAV Receivers and Measurement Error
1 Another viable model would be to use the time difference
of arrival rather than time delay. Using this model the control
The accuracy of the UAV measurements is a
algorithms developed later would remain unchanged since the function of the transmitted radar signal, the signal
observation Jacobians with respect to target parameters are identical strength, the length of the propagation path from
under both models. transmitter to receiver, and so on. To quantify the

ZHAN ET AL.: ADAPTIVE MOBILE SENSOR POSITIONING FOR MULTI-STATIC TARGET TRACKING 121

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 We represent the target state vector at time k as
xk = [x y Vx Vy ]T , and define the vector hk (xk ) by
arranging the time-delay and Doppler equations from
each UAV as

hk (xk ) =[1 N f1 fN ]T

where, as described earlier, N is the number of UAV

Fig. 2. Bistatic radar scenario. receivers. Using a linear discrete-time system and an
additive noise observation model, the system is then
time-delay and Doppler measurement accuracy for written as
UAV i, we use the CRB as derived by [28]:
xk = Ak xk1 + !k1 (6)
2
2Er Er ! !t
CRB1i = (3) zk = hk (xk ) + k (7)
N0 Er + N0 !t t2
where Z where Ak is the state transition matrix, !k1
1
1 2 2 N (0, Qk1 ) and k N (0, Rk ) are, respectively, the
!2 = ! jS(j!)j d!
2 1 process and measurement noise. Both are assumed
Z 1 here to be uncorrelated Gaussian random vectors.
@ s (u)
!t = Im us(u) du As with the standard Kalman filter, the EKF
1 @u
Z can be divided into two stages: the time update and
1
t2 = u2 js(u)j2 du measurement update steps. With the time update, we
1 propagate the targets state estimate and prediction
!i = 2fi covariance matrix as

s(u) and S(j!) are, respectively, the complex envelope x

k = Ak xk1 (8)
of the transmitted signal and its Fourier transform,
P T
k = Ak Pk1 Ak + Qk1 : (9)
Im represents the imaginary part of its argument, Er
is the average received signal power, and N0 =2 is the Once we have the measurements from the sensors, we
spectral density of the white bandpass Gaussian noise update our estimate as
from the receiver on each UAV. Inverting (3) yields
the following explicit expression for the CRB: xk = x
k + Kk (zk hk (xk )) (10)
2
N0 (Er + N0 ) t !t Pk = (I Kk Hk )P (11)
CRBi = : (4) k
2Er2 [! 2 t2 (!t)2 ] !t ! 2
where
The accuracy of the second step, finding the Kk = P T T 1
k Hk (Hk Pk Hk + Rk ) : (12)
position and velocity estimates for the target from the
time-delay and Doppler estimates, are quantified in In the above equations, Hk = (@hk =@x)T is a function
the next section by the EKF. This is done through the of the UAVs positions and headings.
Jacobian Hk (defined subsequently) which maps the The basic idea of this paper is to reduce the size
accuracy of the time-delay and Doppler estimates to of the estimation error by updating the heading of the
the position and velocity of the target. UAVs, and hence Hk . In essence, we are applying an
We use the standard bistatic radar equation [29] outer control loop around the Kalman filter.
to model the received signal power Er [30] based on
the geometry of the bistatic radar scenario depicted in D. System Dynamics
Fig. 2:
E G G 2 The approach presented in the next section is
Er = t t 3 r 2 2b (5)
(4) Rb Ri general enough to accommodate any particular model
for Ak , Qk , or Rk . In the simulation studies presented
where Et is the transmit power; Gt , Gr are antenna
later, we assume a simple constant velocity target
gains at the base station and UAV side, respectively;
model described by the state transition matrix
is the wavelength of the transmitted radio signal;
and b is the radar cross section (RCS) of the target. 2 3
1 0 t 0
6 7
6 0 1 0 t 7
C. Extended Kalman Filter Ak = 6 7 (13)
6 7
The need for the EKF is apparent from the 40 0 1 05
nonlinear measurement equations in (1) and (2). 0 0 0 1

122 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
where t is the time interval between samples. We nonlinearities are accounted for through the Jacobian
also use a simple white process noise model: Hk in the EKF.

Qk = diag(x2 , y2 , V2x , V2y ): (14)

III. HEADING ALGORITHMS
The definition of the measurement noise model
is more critical to the problem at hand, since it will In this section we present two algorithms for
be a function of the UAV positions. Assuming that, finding the optimal UAV heading commands from
given the current target state, the time-delay and the EKF. In the first algorithm, the headings are
Doppler estimates at each UAV are independent from chosen to minimize the weighted trace of the EKF
the estimates made by the rest of the UAV team, prediction error covariance matrix one step into
the expression for the CRB derived above can be the future. In the second, UAV headings are found
used to find the elements of the measurement noise that maximize the amount of new information
covariance. For 1 i N, provided by the measurements at the next step. A
8 heuristic simplification of the second algorithm is also
> CRBi (1, 1) j=i presented that leads to a closed-form solution to the
>
>
< problem and considerable computational savings.
1
Rk (i, j) = E(i j ) = CRBi (1, 2) j = i+N
>
> 2
>
: A. Error Covariance Minimization
0 elsewhere:
The motivation underlying the approach presented
When N < i 2N,
8 2 here is simply to fly the UAVs in directions at
>
> 1 time t = k such that the estimation error at time
>
> CRBiN (2, 2) j =i
>
< 2 t = k + 1 is made as small as possible. To this end,
Rk (i, j) = E(i j ) = we minimize the weighted trace of the one-step
1
>
>
> CRBiN (2, 1) j = iN prediction covariance P k+1 [34]. In practical scenarios,
> 2
>
: the UAV dynamics limit the rate at which its heading
0 elsewhere: may change. Therefore, for each step we restrict the
We assume the ability of the base to pass a UAVs position to an arc defined by the previous
heading command i,k to UAV i. Assuming an heading. Mathematically, the idea is to find the vector
instantaneous response to heading commands, the of heading commands k = [1,k : : : N,k ]T as follows:
UAVs will fly in a straight line with a velocity V
k = arg min tr(Wk P
k+1 )
and heading i during each time interval. The inertial k
(17)
position of the UAV i at time k can be approximated
by ji,k i,k1 j C 8 i = 1:::N

xi,k = xi,k1 + V cos(i,k )t (15) where Wk is a positive semidefinite weighting matrix

and C is the upper bound on the turning rate of the
yi,k = yi,k1 + V sin(i,k )t : (16) UAVs.
Replacing P k+1 in (17) with the expressions in (9)
This approximation will be valid as long as the new
and (11) yields
heading commands are not too different from the
UAVs current heading, and the time steps are short tr(Wk Ak+1 [(P
k)
1
+ HTk R1 1 T
k Hk ] Ak+1 + Wk Qk ):
enough.
(18)
It is known that due to the geometry of a bistatic
radar system, the bistatic radar ambiguity function, Since the dependence of the cost function on the UAV
which is an indicator for how well position and headings is not explicit, it is useful to examine the
velocity can be determined from the received optimization problem in slightly more detail. Equation
waveform, is a nonlinear function of the transmitter, (17) depends on the heading angles through Hk and
receiver, and target positions [3133]. In our system, Rk in (18), which in turn depend on the positions
each UAV calculates only the time-delay and Doppler of the UAVs at time k, which are a function of their
information from the received waveform that is positions at time k 1 and the heading angles chosen
possible to calculate under the assumption that the at time k 1 (as described by the model in (15) and
received signal can be unambiguously associated with (16)). Thus, when we minimize (17) with respect to
the target being tracked. The time-delay and Doppler k , we are minimizing it with respect to the headings
measurements are then passed to the base station for chosen at time k 1 after the measurement at time
processing; the nonlinearities arising in the bistatic k 1, which headings are also the ones in force at
radar ambiguity function are encountered at this step time k before measurement k is taken.
when the velocity and position estimates are extracted When the number of UAVs is not too large, a
from the measured time delay and Doppler. These finite grid search can be used to optimize the cost

ZHAN ET AL.: ADAPTIVE MOBILE SENSOR POSITIONING FOR MULTI-STATIC TARGET TRACKING 123

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
function. For higher dimensional problems, we the covariance measurement update in (11), yielding
employ a Quasi-Newton approach to minimize (18).
P1 1
k = (Pk ) + HTk R1
k Hk (20)
At each iteration the heading is updated by m+1 =
m S1 1
m gm , where Sm and gm are the Hessian matrix where P1k is the information matrix at step k. The
and gradient vector of the cost function evaluated at T 1
term Hk Rk Hk represents the increase in information
m . Due to the complexity involved in calculating the provided by the UAVs, assuming that their headings
Hessian matrix, the Quasi-Newton approach uses an are chosen to yield Hk . If the UAVs provided no
estimate of the Hessian based on the gradient vector. new information at step k, then HTk R1k Hk = 0, and
The gradient with respect to the heading of the ith the information matrix would be equal to its value
UAV at time k is given by after k 1 steps: P1 1
k = (Pk ) . The idea, then, is to
T 1
@tr(Wk P k+1 ) @B 1 T choose Hk so that Hk Rk Hk is maximized in some
= tr Wk Ak+1 B1 B Ak+1 sense. This approach makes sense intuitively: one can
@i,k @i,k
think of this as choosing a set of vectors Hk that lie as
(19) much as possible in the eigenvector space of R1 k with
where the largest eigenvalues (or, equivalently, in the space
B = (P 1
+ HTk R1 of eigenvectors of Rk with the smallest eigenvalues).
k) k Hk
In other words, we try to position the UAVs so that
T the new information that they provide is corrupted the
@B @HTk 1 @HTk 1
= R H + R H least by the measurement noise. Since Hk enters into
@i,k @i,k k k @i,k k k
this term quadratically, an analysis of the impact of
@Rk 1 the UAV headings is much simpler. As discussed in
HTk R1 R H:
k
@i,k k k the next section, this allows for the development of an
efficient algorithm for computing the optimal heading
The problem boils down to calculating @Hk =@i,k and commands.
@Rk =@i,k , where Hk is 2N 4 and Rk is 2N 2N. Under the maximizing information increase (MII)
Minimizing the cost function (17) over the correct approach, the heading commands are found by solving
control variable will work generally for any adaptive
sensor network. The weighting matrix Wk is helpful k = arg max det(HTk R1
k Hk )
k
in that is can be used to place importance on specific (21)
variables of interest. In our problem the magnitude for ji,k i,k1 j C 8 i = 1 : : : N:
position values was much greater than the velocity,
and by adjusting the elements of Wk , we were able to Assuming that the measurement errors across the
place suitable errors on each of these terms. UAVs are uncorrelated, and after regrouping the
time-delay and Doppler equation vector as hk (xk ) =
[1 f1 N fN ]T , the covariance matrix Rk will have
B. Maximizing Information Increase a block diagonal structure. Each block along the
We investigate here an alternative approach diagonal is denoted by Rik , which is equivalent to the
derived from examining an expression for the Fisher CRB matrix discussed in Section IIB. More algebraic
information matrix (or inverse covariance matrix). manipulation leads to the following equality:
Our motivation is to find an algorithm for determining
N
X
the UAV headings that is computationally simpler to 'k = HTk R1
k Hk = HiT i 1 i
k (Rk ) Hk
implement than the error covariance minimization i=1 (22)
(ECM) method of the previous section. As we will HTk = [H1T NT
k : : : Hk ]:
see, the UAV heading parameters enter into the
information matrix update in a much simpler way The submatrices, Hik , which constitute the
that lends itself more easily to approximation. A measurement gradient matrix, HTk , are defined in (34)
closed-form solution based on such an approximation of the Appendix.
is presented in the next section. A computationally By maximizing the new information (21), the
efficient optimization method was found in [9] sensors will position themselves such that the next
for a similar application that used the trace of the measurement contains the most information possible.
covariance matrix, but it was for a much simpler Notice that this cost metric does not necessarily try to
measurement model than assumed here. gain new information that is less known to the system.
Rather than optimizing over the information matrix It simply tries to find the measurements with the
itself, in our approach we choose the UAV headings most information. This metric is perhaps not optimal,
so that the increase in information provided by the but as shown in the results section, it yields good
UAVs at the new locations is maximized [2, 35]. More performance results. Again because of the simple form
precisely, note that the information matrix update can the cost function (21), it allows a closed-form solution
be found by applying the matrix inversion lemma to which is presented in the following section.

124 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
C. Approximate Closed-Form Solution IV. SIMULATION RESULTS

In the Appendix, it is shown that if the relative A. Simulation Setup

positions of the UAVs and target do not dramatically
change over one time step, then the cost function As discussed in Section IIB, the accuracy of the
J(k ) = det 'k in (21) can be approximately decoupled measurements can be adjusted by configuring the
as various transmitted signal parameters. A close look at
J(i,k ) = ki,1 sin(i,k + i,1
k
) (4) reveals that the CRB for time delay and Doppler
shift can be decoupled if the transmitted baseband
k k
+ i,2 sin(2i,k + i,2 ) + i,k + n(i, k) (23) signal is real. In our simulations, we assume a train
of Gaussian-shaped pulses as the transmitted signal, in
k k
where i,m , i,m , i,k , m = 1, 2 are unknown and which case the CRB matrix will assume a diagonal
independent of i,k , n(i, k) accounts for approximation structure. The complex envelope of the signal is
error, and only i,k depends on the headings of the expressed as
other UAVs. Based on this observation, we propose a 2 1=4 X Np
sequential optimization approach that maximizes (23) 1 2 2
s(t) = q e (tkT) =2 (26)
one UAV at a time, while holding the heading of the 2Np + 1 k=Np
others fixed.
Values for ki,1 , i,2 k k
, i,1 k
, and i,2 are found by where T is the interpulse period (IPP), is a
fitting (23) to samples of the criterion function parameter that controls the effective duration of
J(). In particular, treating the approximation error each Gaussian pulse, and 2Np + 1 is the number of
as Gaussian, the following maximum likelihood Gaussian pulses that compose the transmitted signal.
estimators of the unknown parameters can be The choice of the transmitted pulse impacts the
formulated: quality of the time-delay and Doppler estimates,
R 2 ! which must be quantified in order to specify the
k 0 J(t1 ) cos(mt1 )dt1 measurement covariance matrix Rk for each UAV.
i,m = arctan R 2
0 J(t2 ) sin(mt2 )dt2 For the Gaussian pulse train considered here, it can
R 2 ! (24) be shown that !t = 0. Under the assumption that
k
J(t 1 ) sin(mt 1 + i,m )dt1 T 1=, t2 can be approximated as (Np (Np + 1)=3)T2 ,
ki,m = arctan 0R 2 2 :
sin (mt + k
)dt and the CRB for the Doppler estimate can be found
0 2 i,m 2
from (3)(4) to be
In practice, the integrals in (24) are numerically 2 1
evaluated by calculating the cost function at i,k = 1 2Er Er
Var(f f) (t2 )1 :
[0, (1=L)2, : : : , ((L 1)=L)2] and assigning all other 2 N0 Er + N0
UAVs their previous heading at the last iteration. (27)
In the simulations presented later, L = 10 sample Similarly, for the time-delay estimate,
headings were used.
1
Once the amplitudes and phases are estimated, 2Er Er
setting the derivative of the cost function J(i,k ) with Var( ) (! 2 )1
N0 Er + N0
respect to i,k to zero leads to the following quartic Z 1
equation: 1 2 2
2
! = p ! 2 e! =
(2Np + 1) 1
k
(4i,2 k 2
cos i,2 x + ki,1 cos i,1
k k 2
x 2ki,2 cos i,2 ) (28)
0 12
2Np + 1
k 2
k
= (4ki,2 sin i,2 x + ki,1 sin i,1 ) (1 x2 ): (25) B sin !t C
B 2 C d!:
@ !t A
Real-valued solutions x within the unit circle sin
correspond to candidate optimal headings through 2
the relationship cos i,k = x. If multiple candidate While numerical integration can be carried out to
solutions exist within the UAV turning radius, the evaluate this expression, for simplicity we use the
one that yields the largest value for J(i,k ) is chosen. following loose bound to model the time-delay
Real-valued solutions outside the turning radius are estimation accuracy:
evaluated at the nearest boundary point. If there are no 1
real-valued solutions inside the unit circle, J(i,k ) is 2Er Er 2
Var( ) > :
evaluated at the current heading and at the boundaries N0 Er + N0 (2Np + 1)2
of the feasible region, and the heading that results
in the largest value is chosen. The above process is (29)
repeated for each UAV and iterated across the UAV Other signal designs for the transmit signal are
team until the resulting performance gain is negligible. available in the literature [36, 37].

ZHAN ET AL.: ADAPTIVE MOBILE SENSOR POSITIONING FOR MULTI-STATIC TARGET TRACKING 125

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 In the following simulations, we simulate a simple
scenario involving 2 UAVs tracking a moving target.
We assume an average transmit power of 15 W at
the base station (BS) located at the origin, a carrier
frequency of f0 = 10 GHz, number of transmitted
pulses Np = 31, an IPP of T = 1 ms, receive noise
figure F = 5 dB, noise reference temperature T0 =
290 K, and Gaussian pulse parameter = 105 Hz. The
algorithm updates its estimate once every second, and
the heading change constraint C in (17) is set to be
=9.
The initial coordinates of the UAVs were assumed
to be (10, 20) km and (10, 10) km. Both of the Fig. 3. Trajectories of UAVs and target using ECM. A *
UAVs are given initial headings of 0 , and the marker has been placed on each trajectory every 300 s to help
simulation was run for p 1500 s. The UAVs move visualize time evolution.
with constant speed 80 2 m/s. The targets initial
location and speed were (0, 50) km and 100 m/s,
respectively. Its initial heading was randomly created
for each simulation, and in the results presented below
it was about 85 . The targets RCS is assumed
to be b = 100 m2 , and its trajectory is generated
according to (6). For the Kalman filter we set the
initial state estimate to be the true target location
and the covariance matrix associated with the initial
position to be the identity matrix. The process
covariance matrix was set to be Qk = diag(50, 50, 3, 3),
and the time-varying measurement covariance matrix
Rk was generated according to the model described
in Section IID. The antenna gains at the BS side and
the UAV side were assumed to be 40 dB and 30 dB Fig. 4. UAV headings with control using ECM.
respectively.

B. Results
We present results for five simulations involving a
two-agent UAV team tracking a maneuvering target,
one simulation each for the ECM approach of (17),
MII according to (21), the approximate solution of
the previous section, random UAV headings, and
as a comparison, a scenario involving fixed sensor
positions.
In Figs. 35, we plot the performance of the
error covariance minimization approach of (17). The
trajectories of each UAV and the target are shown in
Fig. 3; the starting postions are denoted by crosses Fig. 5. Norm of tracking error using ECM.
(likewise for all subsequent trajectory plots). The
heading angles for each UAV are plotted in Fig. 4. Figs. 911 show the results obtained by the
The UAVs generally tend to follow the target due simplified solution presented in the previous
to the systems desire to increase the SNR (reduce section. As observed, the UAV team behaves almost
path loss), while at the same time spreading out in identically to the MII approach, which shows the
order to improve resolution of the target position and approximation and decoupling work well for this
dynamics. Fig. 5 shows the norm of one run of the scenario regardless of the small number of samples
actual tracking error for both the position and velocity used for the sinusoidal approximation (L = 10 as
estimates. Similar results are shown for the MII mentioned in Section IIIC). The simulation run time
algorithm in Figs. 68. The motion of the UAVs for in Matlab on a 2.16 GHz Intel Core 2 Duo processor
both algorithms is very similar, and both approaches shows that the MIA method is at least 20 times faster
provide similar tracking error performance. than the brute force MII algorithm for this specific

126 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
Fig. 6. Trajectories of UAVs and target using MII. A * marker Fig. 9. Trajectories of UAVs and target using MIA. A *
has been placed on each trajectory every 300 s to help visualize marker has been placed on each trajectory every 300 s to help
time evolution. visualize time evolution.

Fig. 7. UAV headings with control using MII. Fig. 10. UAV headings using MIA.

Fig. 8. Norm of tracking error using MII. Fig. 11. Norm of tracking error using MIA.

case. The tracking error of all of the above algorithms like an antenna element in an adaptive, reconfigurable
is considerably better than that obtained using random array. Clearly, the use of closed-loop heading control
UAV motion (subject to the given turning constraint) significantly reduces the tracking error by arranging
as illustrated in Figs. 1213. Fig. 14 shows the norm the UAVs into a more optimal orientation.
of the actual tracking error for one run of a scenario It is clear from the examples above that there
with sensors fixed at the initial UAV locations. Much are two competing factors that drive the behavior
larger tracking errors are observed in this case. This of the mobile UAV sensors: 1) the desire to reduce
is due to the relatively small sensor aperture and poor the range to the target in order to keep the received
orientation. Here the aperture and orientation refer to signal power from becoming too small, and 2) the
the conglomerate UAV network, where each UAV acts need for the sensors to spread out and provide a

ZHAN ET AL.: ADAPTIVE MOBILE SENSOR POSITIONING FOR MULTI-STATIC TARGET TRACKING 127

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 step ahead. An approach that attempted to predict the
target location at some time farther in the future may
be able to sacrifice a short-term performance gain for
a smaller long-term tracking error. This is the subject
of ongoing research.

V. CONCLUSIONS

In this paper, we have presented several methods

for controlling the movement of mobile sensor
platforms in order to improve target tracking
performance. Two optimization criteria were derived
for the problem, both of which require some type
Fig. 12. Trajectories of UAVs and target with random headings.
A * marker has been placed on each trajectory every 300 s to of search procedure in order to find the desired
help visualize time evolution. solution. Simulations were used to show the benefit
of using closed-loop sensor control for the special
case of an EKF tracking filter. In addition, a
simpler closed-form approach based on one of the
algorithms was also presented, and shown to have
performance similar to that obtained using the optimal
algorithms.

APPENDIX

This Appendix provides justification for the

approximation to the information matrix cost function
in (21).

A. Time-Delay Approximation
Fig. 13. Norm of tracking error with random headings. Recalling the delay expression (1) and the UAV
dynamic model in (15)(16), we can rewrite the
time-delay measurement at the ith UAV at time t = k
as
q q
1
k (i) = xk2 + yk2 + (xk xi,k )2 + (yk yi,k )2
c
q q
1
= xk2 + yk2 + 2
ri,k 2ri,k Vt cos(i,k i,k ) + V2 2t
c

q (30)
ri,k = (xk xi,k1 )2 + (yk yi,k1 )2

yk yi,k1
i,k = arctan :
xk xi,k1

Under the reasonable assumption that the range

Fig. 14. Norm of tracking error with fixed-position sensors.
of each UAV to the target is much larger than the
distance traveled by the UAV in one time step (ri,k
larger aperture for localization. All three of the 2Vt ), we, after droppingp the quadratic term, V2 2t ,
algorithms provide an automated way of balancing use the approximation 1 1 21 for 0 to
these competing objectives. Initially, the UAVs spread show that
apart and move towards the target. As the target q
1
passes, they reverse course, but because of the turning k (i) xk2 + yk2 + ri,k Vt cos(i,k i,k ) :
c
radius constraint and their current headings, they
turn towards each other instead of spreading out. (31)
One reason for this behavior is that the proposed Thus, the time delay is seen to be a sinusoidal
algorithm is somewhat myopic, looking only one time function of the heading angles i,k .

128 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
B. Doppler Shift Approximation assumptions

Inserting the UAV dynamic model (15)(16) into ri,k Vt , ri,k V, ri,k VT (35)
the Doppler measurement model (2) yields q
C where VT = Vx2 + Vy2 , each entry in Hik (except the
fk (i) = 1 f(xi,k1 t Vx xk )V cos i,k
zero terms) can be approximated by a constant plus a
+ (yi,k1 t Vy yk )V sin i,k + C2 g sinusoid with a relatively small amplitude. Assuming
t is small, the second assumption ri,k V includes
Vx xk + Vy yk
+ q (32) the previous cases ri,k Vt and ri,k 2Vt . Then
xk2 + yk2 the two assumptions ri,k V and ri,k VT state,
respectively, that the speed of the UAV and the speed
where of the target must be much smaller than the range of
s the target from the UAV. This ensures that over short
2Vt V2 2 time spans, the bearing angle from the UAV to the
C11 = ri,k 1 cos(i,k i,k ) + 2 t
ri,k ri,k target changes slowly.
From Section IID, we recall that Rik is a constant
and
matrix scaled by N0 (N0 + Er )=2Er2 . When the
C2 = V2 t + Vx (xk xi,k1 ) + Vy (yk yi,k1 ): SNR is not too small, i.e., Er N0 , (Rik )1 can
be approximated as a constant matrix scaled by
Again using the fact that ri,k 2Vt , we have C11 2 SNR. Since Er is inversely proportional to the
ri,k , and hence square of ri,k Vt cos(i,k i,k ), a first-order Taylor
expansion around ri,k can be used to show that the
fk (i) Fi,k Gi,k cos(i,k i,k ) (33) SNR can be approximated as a sinusoid that is again
where offset by a bias term much larger in magnitude. Let
i 1 i
Vx xk + Vy yk 'i,k = HiT
k (Rk ) Hk . Each entry of 'i,k is the product
C
Fi,k = q + 2 of three sinusoids with identical harmonics (i,k ) and
xk2 + yk2 ri,k differing phases plus bias terms that are large relative
V to the magnitude of the sinusoids. Thus the (j, l) entry
Gi,k = [(t Vx + (xk xi,k1 ))2 of 'i,k can be approximated as follows:
ri,k
+ (t Vy + (yk yi,k1 ))2 ]1=2 3
X
'i,k (j, l) aim,k (j, l) sin(m i,k + m (j, l))
and m=1
t Vy + (yk yi,k1 )
i,k = arctan : + bki (j, l) (36)
t Vx + (xk xi,k1 )
As in the case of the time delay, the Doppler shift aip,k (j, l) aiq,k (j, l), if p < q
measurement is also approximately a sinusoidal bki (j, l) aim,k (j, l), 1 m 3:
function of i,k .
In other words, each entry of 'i,k can be characterized
C. Cost Function Approximation as the sum of three harmonics and a large bias,
where the bias is large with respect to the magnitude
Maximizing jAT T 1 1
k+1 Hk Rk Hk Ak+1 j is equivalent to of the three harmonic
P components. Since the cost
maximizing jHTk R1 T 1
k Hk j when both Ak and Hk Rk Hk
function 'k = N i=1 i,k and the determinant of 'k
'
i
are full rank. In (22), each Hk is a 2 4 matrix with involve further products and sums of its elements,
the following structure: coupling between the different UAVs headings
2 i 3 will be introduced. However, due to the relationship
@k @ki
0 0 in (36), the coupling is relatively small compared
6 @xk @yk 7 with the effect of the bias term and the sum of
6 7
Hik = 6 7: (34)
4 @f i @f i @f i @f i 5 first-order sinusoidal terms, where the first-order
k k k k
terms are dependent on only one UAV heading i,k .
@xk @yk @Vx,k @Vy,k
This observation leads to the idea of decoupling
It is easily seen that the gradient of the time delay the N-dimensional optimization problem into N
with respect to the targets position equals the gradient one-dimensional optimization problems. The cost
of the Doppler with respect to the targets velocity function j'k j will produce sinusoidal harmonics up
(i.e., @fki =@Vx,k = @ki =@xk and @fki =@Vy,k = @ki =@yk ). to the twelfth order, but due to the observation in (36),
Using the time-delay and Doppler approximations only harmonics up to the second order are required
given in the previous two sections and under the for sufficient accuracy. Thus, the criterion can be

ZHAN ET AL.: ADAPTIVE MOBILE SENSOR POSITIONING FOR MULTI-STATIC TARGET TRACKING 129

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
written in the following general form: [9] Zhou, K., and Roumeliotis, S.
Optimal motion strategies for range-only distributed target
tracking.
k
J(i,k ) = ki,1 sin(i,k + i,1 )
In Proceedings of the American Control Conference, 2006,
k k 51955200.
+ i,2 sin(2i,k + i,2 ) + i,k + n(i, k)
[10] Martinez, S., and Bullo, F.
(37) Optimal sensor placement and motion coordination for
k k target tracking.
where i,m , i,m , i,k , m = 1, 2 are unknown and
Automatica, 42, 4 (2006), 661668.
independent of i,k , and only i,k depends on the [11] Hernandez, M. L.
headings of the other UAVs. Any errors in this Optimal sensor trajectories in bearings-only tracking.
approximation are absorbed into n(i, k). In P. Svensson and J. Schubert (Eds.), Proceedings of the
The approximation in (37) holds under the Seventh International Conference on Information Fusion,
vol. II, June 2004, 893900.
assumptions in (35). In situations where the range
[12] Oshman, Y., and Davidson, P.
of the target to the UAVs is much larger than the Optimization of observer trajectories for bearings-only
UAVs velocity and the estimated target velocity, this target localization.
approximation will work well, as can be seen from the IEEE Transactions on Aerospace and Electronic Systems,
simulations in Section IVB. As the target gets close 35, 3 (July 1999), 892902.
to the UAVs, higher order effects become important, [13] Grocholsky, B., Makarenko, A., and Durrant-Whyte, H.
Information-theoretic coordinated control of multiple
and the UAV headings become increasingly coupled; sensor platforms.
in such scenarios the optimal heading must be found In Proceedings of the IEEE International Conference on
using a search from the cost function (i.e., (17)). Robotics and Automation, 2003, 15211526.
[14] Maheswararajah, S., and Halgamuge, S.
Mobile sensor management for target tracking.
REFERENCES In Proceedings of the 2nd International Symposium on
Wireless Pervasive Computing, 2007, 506510.
[1] Ousingsawat, J., and Campbell, M. [15] Baras, A., and Bensoussan, J. S.
Establishing trajectories for multi-vehicle reconnaissance. Sensor scheduling problems.
In Proceedings of the AIAA Guidance, Navigation, and In Proceedings of the IEEE Conference on Decision and
Control Conference, Aug. 2004, 21882199. Control, vol. 3, Dec. 1988, 23422345.
[2] Ousingsawat, J., and Campbell, M. [16] Gupta, V., Chung, T., Hassibi, B., and Murray, R.
Optimal cooperative reconnaissance using multiple On a stochastic sensor selection algorithm with
vehicles. applications in sensor scheduling and sensor coverage.
AIAA Journal of Guidance, Control, and Dynamics, 30, 1 Automatica, 42, 2 (2006), 251260.
(2007), 122132. [17] He, Y., and Chong, K.
[3] Hernandez, M., Kirubarajan, T., and Bar-Shalom, Y. Sensor scheduling for target tracking in sensor networks.
Multisensor resource deployment using posterior In Proceedings of the IEEE Conference on Decision and
Cramer-Rao bounds. Control, Dec. 2004, 743748.
IEEE Transactions on Aerospace and Electronic Systems, [18] Kaplan, L.
40, 2 (2004), 399416. Global node selection for localization in a distributed
sensor network.
[4] Sinha, A., Kirubarajan, T., and Bar-Shalom, Y.
IEEE Transactions on Aerospace and Electronic Systems,
Autonomous ground target tracking by multiple
42, 1 (Jan. 2006), 113135.
cooperative UAVs.
In Proceedings of the IEEE Aerospace Conference, 2005, [19] Kaplan, L.
19. Local node selection for localization in a distributed
sensor network.
[5] Kovacina, M., Palmer, D., Yang, G., and Vaidyanathan, R.
IEEE Transactions on Aerospace and Electronic Systems,
Multi-agent control algorithms for chemical cloud
42, 1 (Jan. 2006), 136146.
detection and mapping using unmanned air vehicles.
[20] Mehrotra, K., Mohan, C., Niu, R., and Varshney, P.
In Proceedings of the IEEE/RSJ International Conference
Temporally staggered sensors in multi-sensor target
on Intelligent Robots and System, vol. 3, Sept. 30Oct. 5,
tracking systems.
2002, 27822788.
IEEE Transactions on Aerospace and Electronic Systems,
[6] Keller, D. R., Moon, T. K., and Gunther, J. H. 41, 3 (July 2005), 794808.
Narrowband source localization from a moving array of [21] Moges, M., and Robertazzi, T.
sensors. Wireless sensor networks: Scheduling for measurement
In Proceedings of the 40th Asilomar Conference Signals, and data reporting.
Systems, and Computers, 2006, 22852289. IEEE Transactions on Aerospace and Electronic Systems,
[7] Toussaint, G. J., De Lima, P., and Pack, D. J. 42, 1 (Jan. 2006), 327340.
Localizing RF targets with cooperative unmanned aerial [22] Erdinc, O., Willett, P., and Coraluppi, S.
vehicles. Multistatic sensor placement: A tracking approach.
In Proceedings of the American Control Conference, July In Proceedings of the 9th International Conference on
913, 2007, 59285933. Information Fusion, July 2006, 18.
[8] Pack, D., York, G., and Fierro, R. [23] Punithakumar, K., Kirabarajan, T., and Hernandez, M.
Information-based cooperative control for multiple Multisensor deployment using PCRLBs incorporating
unmanned aerial vehicles. sensor deployment and motion uncertainties.
In Proceedings of the IEEE International Conference on IEEE Transactions on Aerospace and Electronic Systems,
Networking, Sensing and Control, 2006, 446450. 42, 4 (Oct. 2006), 14741485.

130 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
[24] Keller, D., Moon, T., and Gunther, J. [26] Beard, R., Kingston, D., Quigley, M., Snyder, D.,
Source localization from moving arrays of sensors. Christiansen, R., Johnson, W., McLain, T., and Goodrich, M.
In Proceedings of the 39th Asilomar Conference on Signals, Autonomous vehicle technologies for small fixed-wing
Systems, and Computers, 2005, 452456. UAVs.
[25] Gu, G., Chandler, P., Schumacher, C., Sparks, A., and Journal of Aerospace Computing, Information, and
Prachter, M. Communication, 2, 1 (2005), 92108.
Optimal cooperative sensing using a team of UAVs.
IEEE Transactions on Aerospace and Electronic Systems,
42, 4 (2006), 14461458.

Pengcheng Zhan received the B.S. degree in electrical engineering from Tsinghua
University, Beijing, China in 2002, and the Ph.D. degree in electrical engineering
from Brigham Young University, Provo, UT, in 2007.
He is employed at ArrayComm LLC, San Jose, CA, where he works on
multi-antenna signal processing for wireless communications.

ZHAN ET AL.: ADAPTIVE MOBILE SENSOR POSITIONING FOR MULTI-STATIC TARGET TRACKING 131

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.
 David W. Casbeer received the B.S. degree in electrical engineering from
Brigham Young University, Provo, UT, in 2004. He is currently a Ph.D. candidate
in electrical engineering at Brigham Young University.
Since 2003, he has worked as a research assistant in the MAGICC Lab,
Brigham Young University, where he has worked on multiagent estimation and
control problems. He visited the Australian Centre for Field Robotics, University
of Sydney, Sydney, Australia in the summer of 2007 where he investigated the
use of consensus filters in decentralized data fusion.

A. Lee Swindlehurst (M84SM99F04) received the B.S., summa cum

laude, and M.S. degrees in electrical engineering from Brigham Young University,
Provo, UT, in 1985 and 1986, respectively, and the Ph.D. degree in electrical
engineering from Stanford University, Stanford, CA, in 1991.
From 19861990, he was employed at ESL, Inc., of Sunnyvale, CA, where
he was involved in the design of algorithms and architectures for several radar
and sonar signal processing systems. He was on the faculty of the Department
of Electrical and Computer Engineering at Brigham Young University from
19902007, where he was a full professor and served as department chair from
20032006. During 19961997, he held a joint appointment as a visiting scholar
at both Uppsala University, Uppsala, Sweden, and at the Royal Institute of
Technology, Stockholm, Sweden. From 20062007, he was on leave working as
Vice President of Research for ArrayComm LLC in San Jose, CA. He is currently
a Professor of Electrical Engineering and Computer Science at the University of
California at Irvine. His research interests include sensor array signal processing
for radar and wireless communications, detection and estimation theory, and
system identification, and he has over 160 publications in these areas.
Dr. Swindlehurst is past Secretary of the IEEE Signal Processing Society. He
is currently serving as Editor-in-Chief of the IEEE Journal of Selected Topics in
Signal Processing, as a member of the Editorial Boards for the EURASIP Journal
on Wireless Communications and Networking and the IEEE Signal Processing
Magazine, and is a past associate editor for the IEEE Transactions on Signal
Processing. He is a recipient of several paper awards: the 2000 IEEE W. R. G.
Baker Prize Paper Award, the 2006 IEEE Signal Processing Societys Best Paper
Award, the 2006 IEEE Communications Society Stephen O. Rice Prize in the
Field of Communication Theory, and is coauthor of a paper that received the
IEEE Signal Processing Society Young Author Best Paper Award in 2001.

132 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:534 UTC from IE Xplore. Restricon aply.