Signal Processing 115 (2015) 144150

Contents lists available at ScienceDirect

Signal Processing
journal homepage: www.elsevier.com/locate/sigpro

Fast communication

Weighted least squares algorithm for target localization

in distributed MIMO radar
Martin Einemo a,b, Hing Cheung So a,,n
a
b

Department of Electronic Engineering, City University of Hong Kong, Hong Kong, China
Centre for Mathematical Sciences, Lund University, Sweden

a r t i c l e in f o

abstract

Article history:
Received 19 July 2014
Received in revised form
27 March 2015
Accepted 3 April 2015
Available online 11 April 2015

In this paper, we address the problem of locating a target using multiple-input multipleoutput (MIMO) radar with widely separated antennas. Through linearizing the bistatic
range measurements, which correspond to the sum of transmitter-to-target and target-toreceiver distances, a quadratically constrained quadratic program (QCQP) for target
localization is formulated. The solution of the QCQP is proved to be an unbiased position
estimate whose variance equals the CramrRao lower bound. A weighted least squares
algorithm is developed to realize the QCQP. Simulation results are included to demonstrate the high accuracy of the proposed MIMO radar positioning approach.
& 2015 Elsevier B.V. All rights reserved.

Keywords:
Multiple-input multiple-output (MIMO)
radar
Target localization
Bistatic range
Weighted least squares

1. Introduction
Target localization [1] has been one of the central
problems in many fields such as radar [2], sonar [3],
telecommunications [4], mobile communications [5], sensor networks [6] as well as humancomputer interaction
[7]. Recently, this topic has also received considerable
interest in multiple-input multiple-output (MIMO) radar
[813]. Unlike the conventional phased-array radar which
deals with a single waveform, MIMO radar employs multiple antennas to transmit and receive different waveforms
and process the received signals. The MIMO radar with
colocated antennas utilizes waveform diversity while that
with widely separated antennas provides spatial diversity,
and both of them are superior to its phased-array counterpart in many aspects. In this work, we address target
localization using the latter architecture.

EURASIP Member.
Corresponding author.
E-mail address: hcso@ee.cityu.edu.hk (H.C. So).

http://dx.doi.org/10.1016/j.sigpro.2015.04.004
0165-1684/& 2015 Elsevier B.V. All rights reserved.

Generally speaking, MIMO radar positioning with distributed antennas can be classified as direct and indirect
approaches. In the former, the target location is estimated
via processing the observed receiver outputs directly, and
representative examples include the maximum likelihood
(ML) [9,10] and sparse recovery [11] methods. On the other
hand, there are two steps in the indirect approach. First, the
time delay information, which corresponds to the sum of the
signal propagation time from a transmit antenna to the
target and that from the target to a receive antenna, is
estimated from the received data. In the second step, we
multiply these delays by the known signal propagation
speed to yield the bistatic range estimates for constructing
a set of elliptic equations from which the source position can
be determined. Godrich et al. [9] have proposed a best linear
unbiased estimator (BLUE) by linearizing the elliptic equations via Taylor series expansion [14] while a different
linearization scheme is devised in [12], which results in
solving two sets of linear equations. We refer the latter
solution to as the combined linear least squares (CLLS)
method. Although the BLUE is superior to the CLLS solution

M. Einemo, H.C. So / Signal Processing 115 (2015) 144150

and can attain ML performance, an initial position estimate

which is sufficiently close to the source location is required.
Generally speaking, the direct methods are more computationally demanding than the indirect approach. For example,
according to [9], when the ML estimator is realized by a grid
search, a significant computational effort is involved. The
complexity of the sparse modeling approach [11] is also high
because convex optimization is involved. Inspired by the
linearization approach of Chan and Ho [15], we devise an
accurate indirect positioning approach for distributed MIMO
radar where no algorithm initialization is needed. Our
contributions can be summarized as (i) the problem of
solving the elliptic equations constructed from the bistatic
range estimates is converted to a quadratically constrained
quadratic program (QCQP); (ii) we have proved that under
sufficiently small noise conditions, the solution of the QCQP
is an unbiased estimate of the target location and its variance
is equal to the optimality benchmark, namely, the Cramr
Rao lower bound (CRLB); and (iii) analogous to the relaxation
technique in [15], we develop a weighted least squares
(WLS) estimator for solving the QCQP.
The rest of the paper is organized as follows. In Section
2, the problem of distributed MIMO radar positioning is
formulated. In Section 3, we first convert the target
localization formulation to a QCQP. Bias and mean square
error (MSE) of the estimate of the QCQP are also analyzed.
The WLS algorithm for realizing the QCQP is then developed. Simulation results are presented in Section 4 to
evaluate the localization accuracy of the proposed
approach by comparing with the CLLS method and the
CRLB. Finally, conclusions are drawn in Section 5.
The notation is introduced as follows. Scalars, vectors
and matrices are denoted by italic, bold lower-case and
bold upper-case symbols, respectively. An estimate of a is
^ The ith element of a and (i,j) entry of A are
denoted by a.
represented as ai and Ai;j , respectively. Moreover, Ai:j;k:l
contains entries in the intersection of the ith to the jth
rows and the kth to the lth columns. The Euclidian norm of
a is denoted by a2 . The T and  1 denote the matrix
transpose and inverse, respectively, while E is the expectation operator. The diaga1 ; a2 ; ; ak is a diagonal matrix
with diagonal elements a1 ; a2 ; ; ak . The 0ij and Ii represent the i  j zero matrix and i  i identity matrix,
respectively.
2. Problem formulation
We consider finding the position of a target, denoted by
x x yT , using a distributed MIMO radar system with M
transmit and N receive antennas whose coordinates,
xtm xtm ytm T , m 1; 2; ; M, and xrn xrn yrn T , n
1; 2; ; N, respectively, are known. Note that although we
assume two-dimensional positioning, extension to the threedimensional case is straightforward. Suppose the transmit
antennas send a set of orthogonal waveforms with center
frequency fc and bandwidth f . Let sm(t) be the low-pass
equivalent of the emitted signal from the mth transmitter.
The transmitted waveforms are reflected by the target and
then collected at the receive antennas. In the case of
coherent processing where the receivers are phase-synchronized, the signal measured at the nth antenna, denoted by

145

zn(t), is [9,12]
zn t

M
X

m;n expf j2 f c m;n gsm t  m;n wn t:

m1

Here, m;n is the amplitude corresponding to the path from

the mth transmitter to the nth receiver, m;n is the sum of the
signal propagation time from the mth transmitter to the
target and the time from the target to the nth receiver,
denoted by tm and rn , respectively, and wn(t) is the zeromean complex Gaussian noise which is temporally and
2
spatially white with variance w. That is, m;n can be
decomposed as

m;n tm rn ;

m 1; 2; ; M; n 1; 2; ; N

Denoting the wave propagation speed by c,

also be expressed as
q
1
1
tm xtm  x2
xtm  x2 ytm  y2
c
c
1
c

rn xrn  x2

1
c

2
t
m

and can

q
xrn  x2 yrn  y2 :

r
n

3
4

For non-coherent processing model where the receive

antennas are not phase-synchronized, the observed signal
becomes [9,12]
zn t

M
X

m;n sm t  m;n wn t:

m1

In the direct approach, x is determined directly from (1)

or (5). In this work, our focus is on the indirect approach
where we assume that fm;n g have been estimated from
the received signals using the ML method. By multiplying
the estimates by c, we obtain the bistatic range measurements, denoted by fr m;n g, which are modeled as
r m;n Rtm Rrn m;n ;

m 1; 2; ; M; n 1; 2; ; N

where
Rtm xtm  x2

Rrn xrn x2 :

t
Rm

8
r
Rn

and
are the unknown ranges between the
Here,
target and the mth transmitter and the nth receiver,
respectively, while m;n is the zero-mean Gaussian distributed range estimation error, whose power is inversely
proportional to jm;n j2 . For presentation simplicity, we
assume that jm;n j jj for all m and n. In the coherent
case, the covariance matrix of m;n , denoted by C , is [9]
C

c2 2w
8 2 f c jj2
2

IMN :

While for non-coherent processing, the corresponding

covariance matrix is given by [9]
C

c2 2w

8 f 2 jj2
2

IMN :

10

In both cases, C is a scaled identity matrix, indicating that

m;n is zero-mean white Gaussian process and we denote its

variance as 2 . It is worth noting that our algorithm devel-

opment and analysis can be straightforwardly extended to

the case when each m;n has distinct power, that is, C is a

146

M. Einemo, H.C. So / Signal Processing 115 (2015) 144150

diagonal matrix. Given the MN bistatic range measurements

as well as fxtm g and fxrn g, the task is to find x.

W Efqq g

1

0
1
1
T
T
TA
@
 2 diag w ; w ; ; w
|{z}
4

24

N terms

3. Algorithm development

where
"

3.1. Reformulation as QCQP

Following the idea of [15], we first reorganize (6) as

r m;n  Rrn Rtm m;n ;

m 1; 2; ; M; n 1; 2; ; N:

11

Squaring both sides of (11) and using (7)(8), we obtain

r 2m;n xrn 2  xtm 2 yrn 2  ytm 2
2xrn  xtm x 2yrn  ytm y 2r m;n Rrn 2m;n 2Rtm m;n :
12
It is clear that the MN equations of (12) are linear in x, y
and fRrn g. Let

x y Rr1 RrN T :

13

We can express (12) in a compact form as

A q b
where
2
B1
6
6 B2
A6
6
4
BN

14
3

1
7
2 7
7

15

7
5

q qT1 qT2 qTN T

T

16

b b1 b2 bN T :

17
M2

The submatrix Bk A R
has the form of
2 r
3
xk  xt1 yrk  yt1
6 r
7
6 xk  xt2 yrk  yt2 7
7
Bk 26
6
7

4
5
xrk  xtM yrk  ytM

3
r 2 xr1 2 xtk 2 yr1 2  ytk 2
6 1;k
7
6 r 2 xr 2 xt 2 yr 2  yt 2 7
6 2;k
7
2
2
k
k
bk 6
7:
6
7

4
5
2
r 2
t 2
r 2
t 2
r M;k xM  xk yM  yk

18

min J

25

s:t:

Rrn 2 xrn  x22 ; n 1; 2; ; N

26

which is a QCQP because the objective function J and all

N constraints are of quadratic forms. It is worth mentioning that an equivalent form of (26) is to reorganize (6) by
considering fRtm g as unknowns to be determined, which
results in M constraints given by (7). Apparently, it is
simpler to solve the QCQP with less constraints in the
algorithm realization. Without loss of generality, we
assume N rM, that is, the number of receive antennas is
equal to or less than the number of transmit antennas.
3.2. Bias and MSE analysis of QCQP estimate
To analyze the solution of the QCQP, our first step is to
convert (26) to an unconstrained optimization problem,
which is achieved by substituting the N constraints into
the objective function. The resultant cost function, denoted
by Jx, is not characterized by fRrn g but is highly nonlinear.
The position estimate of (26) is now written as
27

where
q32
2
M X
N
xt  x2 ytm  y2  rm;n xrn  x2 yrn  y2
1 X
4 m
5 :
4 2 m 1 n 1
Rtm

28

20

Applying the bias and MSE analysis for unconstrained

optimization problems [17], we have (see Appendix)

  2
  1 
Jx
Jx
^ x  E
021
E
29
Efxg
xxT
x
and

21

22



Cx^ Ef x^  x x^ xT g


 )  2
  2
  1 (

  1
Jx
Jx
Jx T
Jx
E
 E
E
xxT
x
x
xxT
30
where

MNMN

is a symmetric weighting matrix.

where W A R
When fm;n g are sufficiently small such that jRtm j c jm;n j,
we can write

and the optimal weight is [16]

RtM 2

#T

19

Based on the linear model of (14), we construct a WLS cost

function [16], denoted by J:

qk  2Rt1 1;k Rt2 2;k RtM M;k T

To fully utilize the available information, we notice the

relationship between the parameters in , which is given
by (8). As a result, we propose to find the source location
via solving the following constrained optimization problem:

J x

J A bT WA  b

Rt1 2 Rt2 2

One the other hand, the components in q and b are

qk 21;k 2Rt1 1;k 22;k 2Rt2 2;k 2M;k 2RtM M;k T

x^ argmin Jx

while k A RMNN has zero entries except

k 1:M;k 2r 1;k r 2;k r M;k T :

23

1
Cx
^ 1;1

M X
N
1 X

2 m 1 n 1

1
1
Cx
^ 1;2 Cx^ 2;1

xtm x xrn x
r
Rn
Rtm

M X
N
1 X

2 m 1 n 1

!2
31

M. Einemo, H.C. So / Signal Processing 115 (2015) 144150

!
!
xtm  x xrn  x
ytm  y yrn  y

Rrn
Rrn
Rtm
Rtm
1
Cx
^ 2;2

M X
N
1 X

2 m 1 n 1

!2
ytm  y yrn  y
r
:
Rn
Rtm

Note that (29) and (30) are accurate when the noise is
small and J 2 x=xxT is sufficiently smooth around the
1
true value. In the Appendix, we have also proved that Cx
^
is equal to the Fisher information matrix (FIM) for distributed MIMO radar positioning. That is to say, the MSE of
x^ attains the CRLB.
3.3. WLS algorithm
As directly solving the nonconvex QCQP is a computationally difficult problem, our algorithm is to relax (26) by
using the WLS technique. Similar to [15], we first compute
the WLS solution of by minimizing J and then exploit
the constraints in (26) to improve the position estimate
using WLS again.
The WLS estimate of based on minimizing the quadratic
function J is straightforwardly computed as [16]
34

fRtm g

Note that since

are unknown, we first use the least
squares solution of AT A  1 AT b for their estimation prior to
h
iT
in (34) is the
constructing W. It is clear that ^ 1 ^ 2
estimate of x. To enhance the localization accuracy, we make
use of the last N elements of ^ according to the relation of (8)
as follows. Under small noise conditions, x^ of (34) will be
sufficiently close to x, we then have

2
h
i
xrn  ^ 1  xrn  x2  2xrn x xrn  ^ 1 xrn  x
35


yrn ^ 2

 yrn  y2  2yrn y

h
i
yrn  ^ 2  yrn  y
36

h i2
^

n2

h i
 2
 Rrn  2Rrn ^

n2

 Rrn ;

n 1; 2; ; N:

37

As it is not feasible to relate (35) and (36) with all N

equations of (37) simultaneously, we resort to do it in a
separate manner. Denote the estimates of x and y associated
n
n
with the nth equation of (37) by x^ and y^ , respectively. In
matrix form, we have
n

Gzn pn

38

where

h i 2 
h i 2 h i2
n
yrn  ^
^
h xrn  ^
1

6
G40
1

n2

T
39

7
15

40

h
iT
n
n
zn xrn  x^ 2 yrn  y^ 2

h
i3
xr  x xrn  ^ 1  xrn  x
6 n
h
i7
6
^   yr y 7
r
r
7:
pn 26

y
y


y
2
n
n
6 n
7
4
5
r ^
r
Rn n 2  Rn
2

32

33

^ argmin J AT WA  1 AT Wb:

147

41

42

Employing the fact that the covariance matrix for ^ is

AT WA  1 [16], the covariance for pn , denoted Cpn , is
determined as follows. Let
Cp 4diagx; y; Rr1 ; ; RrN AT WA  1 diagx; y; Rr1 ; ; RrN :
43
Then, Cpn has the form of
"
#
Cp 1:2;1:2
Cp 1:2;n 2
Cpn
:
Cp n 2;1:2 Cp n 2;n 2

44

Similar to (34), the optimal WLS solution of (38) is

n
n
z^ GT Cpn  1 G  1 GT Cpn  1 h :

45

are not available, we employ ^ 1

Note that since x, y and
^
and 2 as well as (8) to construct Cpn in practice. From (45),
we can deduce
q
q
n
n
n
n
x^ xrn z^ 1 or x^ xrn  z^ 1
q
q
n
n
n
n
y^ yrn z^ 2 or y^ yrn  z^ 2 :
46
fRrn g

n
n
To find the appropriate pair, we choose the x^ ; y^ which
is closest in the 2-norm sense (34). Assuming that the N sets
n
n
of fx^ g and fy^ g are independent, we obtain the final
solution using the WLS again:
" PN
#
PN
^ n
^ n


n 1 x =Cpn 1;1
n 1 y =Cpn 2;2
:
47
x^ y^
P
PN
1= N
n 1 Cpn 1;1 1=
n 1 Cpn 2;2

4. Numerical examples
Computer simulations are conducted to evaluate the
MSE performance of the proposed approach by comparing
with the CLLS method [12] and CRLB. The solution based
on (34), which is referred to as WLS (one-step), is also
examined. The target position is fixed at 500; 500 km.
We scale fm;n g to produce different noise conditions. All
results provided are averages of 1000 independent runs.
In the first test, we consider the geometry of the MIMO
radar system depicted in Fig. 1, which aligns with [12].
There are M 4 transmit antennas and N 3 receive
antennas where the position of one transmitter is 0; 0
while the remaining antennas are located asymmetrically
on the circumference of a circle with radius of 120 km.
Fig. 2 plots the MSE versus 2 . We can see that the
performance of the proposed method attains CRLB when
the noise power is sufficiently small, that is, 2 r108 . Even
the first-step WLS estimate of (34), although it only
provides suboptimal estimation performance, is significantly superior to the CLLS scheme.
In the second test, there are M 3 transmit antennas
and N 2 receive antennas where we remove the top
transmit and receive antennas in Fig. 1. The results are
plotted in Fig. 3. Same findings are observed although the
MSE is now higher because of a smaller number of antenna

148

M. Einemo, H.C. So / Signal Processing 115 (2015) 144150

5

x 10
1

25

10

Transmit antenna
Receive antenna

WLS (one-step)
WLS
CRLB

20

10

1015

MSE (m2)

yaxis (m)

0.5

1010

10
0.5

100
1
1.5

0.5

0.5

1.5

xaxis (m)

x 10

1020

20

40

60

80

100

120

140

160

10log10(2)

Fig. 3. MSE versus 2 when M 3 and N 2.

algorithm development requires jRtm j c jm;n j, which is valid

in normal operating conditions.

CLLS
WLS (one-step)
WLS
CRLB

Appendix A

15

10

First, we will provide the detailed derivations for (29)

(33). We start with evaluating Jx=x. Let

MSE (m )

- 20

Fig. 1. Geometry of distributed MIMO radar system with M 4 and N 3.

1025

10-5
-40

10

10

dm;n Rtm Rrn :

A:1

According to (28), the two elements in Jx=x are


M X
N



Jx
1 X
 2r m;n xrn  x
2xrn  2xtm rm;n  dm;n
2
r
x
2 m 1 n 1
Rn

105

10 0

A:2
-5

10

-40

-20

20

40

60

80

100

120

140

160

10log ( )

Fig. 2. MSE versus


M X
N



Jx
1 X
2rm;n yrn y
2
2yrn  2ytm rm;n  dm;n :
r
y
2 m 1 n 1
Rn

A:3

when M 4 and N 3.

elements. Note that the CLLS method cannot work in this

case because it requires M Z 3 and N Z 3.
5. Conclusion
Given the time delay measurements of a distributed
MIMO radar system, a simple and accurate approach for
target localization is devised. We first reformulate the
problem as a QCQP and then solve it in a relaxed manner
according to the WLS technique. Both theoretical calculations
and simulation results confirm the optimality of the proposed positioning algorithm. Nevertheless, the signal model
of (6) assumes line-of-sight (LOS) propagation, that is, there
are direct propagation paths from all transmit antennas to
the target as well as from the target to all receive antennas,
and absence of multipaths. Note that when there are multipaths, delay estimation algorithm [18] can be applied to
extract the propagation time which corresponds to the LOS
path. It is also assumed that the positions of the transmitters
and receivers are perfectly known, but this may not be
fulfilled in some practical scenarios [19,20]. Finally, our

Since Efr m;n g dm;n , we have



Jx
021
E
x

A:4

For 2 Jx=xxT , its (1,1) entry is

2 Jx
2 x
"
!
M X
N

xtm  x2 1
2 X
1 xrn  x2
2
r m;n  dm;n



r
m 1 n 1
Rtm Rn
Rtm 2
Rrn 2
!2 3
xtm  x xrn  x 5
:
A:5
r

Rn
Rtm
Using Efr m;n g dm;n again and (A.5), we get
!2
 2

M X
N
Jx
2 X
xtm  x xrn  x
E

:
2 x
2 m 1 n 1 Rtm
Rrn

A:6

In a similar way, we have

!2
 2

M X
N
Jx
2 X
ytm y yrn y
E
2
r
2 y
m 1 n 1 Rtm
Rn

A:7

M. Einemo, H.C. So / Signal Processing 115 (2015) 144150

 2

 2

M X
N
Jx
Jx
2 X
E
2
xy
yx
m 1 n 1
!
!
xtm x xrn  x
ytm  y yrn  y
:
r
r
Rn
Rn
Rtm
Rtm

"


r m;n  dm;n
A:8

From (A.6)(A.8), we deduce that Ef2 Jx=xxT g is nonsingular. Together with the result of (A.4), (29) is proved.
To determine the covariance, we also need the expression of Jx=xJx=xT . Utilizing (A.2), its (1,1)
entry is computed as


M X
N
X
Jx 2
1
r m;n xrn  x 2

r m;n  dm;n 2 :
r
t
4
4
x
Rn
m 1 n 1 Rm
A:9
Taking the expectation on (A.9) and using (23) that the
variance of r m;n is approximately equal to 4Rtm 2 2 , yields
(

!2
 )
M X
N
Jx 2
4 X
xtm  x xrn x
2
E
r
:
A:10
x
m 1 n 1 Rtm
Rn
Similarly, we obtain



M X
N
Jx
Jx
4 X
2
E
x
y
m 1 n 1
!
!
xtm  x xrn  x
ytm  y yrn  y
r
r
Rn
Rn
Rtm
Rtm

!2
 )
M X
N
Jx 2
4 X
ytm  y yrn  y

:
2
y
m 1 n 1 Rtm
Rrn

A:17
Since Efr m;n g dm;n , we then have
!2
 2

M X
N
ln pfr m;n g; x
1 X
xtm x xrn x
2
F1;1 E
r
:
m 1 n 1 Rtm
2 x
Rn

A:18
In a similar manner, we obtain
F1;2 F2;1

M X
N
1 X

2 m 1 n 1

!
!
xtm  x xrn  x
ytm  y yrn  y

Rrn
Rrn
Rtm
Rtm

A:19
M X
N
1 X

2 m 1 n 1

!2
ytm  y yrn  y

:
Rrn
Rtm

A:20

The CRLB is then determined from F  1 .

A:11

A:12

Finally, (31)(33) are obtained.

Now we derive the CRLB for x as follows. To determine
it, we first need the probability density function of fr m;n g,
which has the form of a multivariate normal distribution:
(
)
M X
N


1
1 X
2
:
p fr m;n g; x
exp

r

d

m;n
m;n
2 2 m 1 n 1
2 MN=2 MN

A:14
The CRLB is given by the diagonal elements of the inverse
of the corresponding Fisher information matrix, denoted
by F:
o
n2
o3
2 n2
pfr m;n g;x
m;n g;x
E ln xy
E ln pfr
2x

6
n2
o7
A:15
F  4 n2 ln pfr g;xo
5:
m;n
m;n g;x
E ln pfr
E
yx
2 y
To compute (A.15), we first take the natural logarithm on
(A.14) to obtain
"
#
M X
N



MN 
1 X
2
2
ln p fr m;n g; x 
r m;n  dm;n :
ln 2 
2
2 2 m 1 n 1
A:16

M X
N
2 ln pfr m;n g; x
1 X


2 x
2 m 1 n 1

xtm  x2
1
 t
t
xm  x2 ytm  y2=3 Rm

#
!2 1
1
xrn  x2
xtm x xrn x A
:
r

 r
Rn xrn  x2 yrn  y2=3
Rn
Rtm

F2;2

From (A.6)(A.8) and (A.10)(A.12), we can conclude that

(

 )
 2

Jx
Jx T
Jx
2E
E
:
A:13
x
x
xxT

The second derivative of (A.16) with respect to x is

149

References
[1] H.C. So, Source localization: algorithms and analysis, in: S.A. Zekavat,
M. Buehrer (Eds.), Handbook of Position Location: Theory, Practice
and Advances, Wiley-IEEE Press, Hoboken, N.J., 2011. (Chapter 2).
[2] B.K. Chalise, Y.D. Zhang, M.G. Amin, B. Himed, Target localization in a
multi-static passive radar system through convex optimization,
Signal Process. 102 (2014) 207215.
[3] G.C. Carter (Ed.), Coherence and Time Delay Estimation: An Applied
Tutorial for Research, Development, Test, and Evaluation Engineers,
IEEE, New York, 1993.
[4] Y. Huang, J. Benesty (Eds.), Audio Signal Processing for NextGeneration Multimedia Communication Systems, Kluwer Academic
Publishers, Boston, 2004.
[5] J.J. Caffery Jr., Wireless Location in CDMA Cellular Radio Systems,
Kluwer Academic Publishers, Boston, 2000.
[6] N. Patwari, J.N. Ash, S. Kyperountas, A.O. Hero III, R.L. Moses, N.
S. Correal, Locating the nodes: cooperative localization in wireless
sensor networks, IEEE Signal Process. Mag. (July) (2005) 5469.
[7] T.K.C. Lo, H.C. So, Theoretical study of impact localization limits in
tangible acoustic interface applications, IET Electron. Lett. 45 (July
(16)) (2009) 856858.
[8] M. Jin, G. Liao, J. Li, Joint DOD and DOA estimation for bistatic MIMO
radar, Signal Process. 89 (2009) 244251.
[9] H. Godrich, A.M. Haimovich, R.S. Blum, Target localization accuracy
gain in MIMO radar-based systems, IEEE Trans. Inf. Theory 56 (June
(6)) (2010) 27832803.
[10] O. Bar-Shalom, A.J. Weiss, Direct positioning of stationary targets
using MIMO radar, Signal Process. 91 (2011) 23452358.
[11] S. Gogineni, A. Nehorai, Target estimation using sparse modeling for
distributed MIMO radar, IEEE Trans. Signal Process. 59 (November
(11)) (2011) 53155325.
[12] M. Dianat, M.R. Taban, J. Dianat, V. Sedighi, Target localization using
least squares estimation for MIMO radars with widely separated
antennas, IEEE Trans. Aerosp. Electron. Syst. 49 (October (4)) (2013)
27302741.
[13] B. Tang, J. Tang, Y. Zhang, Z. Zheng, Maximum likelihood estimation
of DOD and DOA for bistatic MIMO radar, Signal Process. 93 (2013)
13491357.
[14] W.H. Foy, Position location solutions by Taylor-series estimation,
IEEE Trans. Aerosp. Electron. Syst. 12 (March (2)) (1976) 187194.
[15] Y.T. Chan, K.C. Ho, A simple and efficient estimator for hyperbolic
location, IEEE Trans. Signal Process. 42 (August (8)) (1994)
19051915.
[16] S.M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory, Prentice-Hall, Englewood Cliffs, NJ, 1993.

150

M. Einemo, H.C. So / Signal Processing 115 (2015) 144150

[17] H.C. So, Y.T. Chan, K.C. Ho, Y. Chen, Simple formulas for bias and
mean square error computation, IEEE Signal Process. Mag. 30 (July
(4)) (2013) 162165.
[18] J.-J. Fuchs, Multipath time-delay detection and estimation, IEEE
Trans. Signal Process. 47 (February (1)) (1999) 237243.
[19] K.W.K. Lui, W. Ma, H.C. So, F.K.W. Chan, Semi-definite programming
algorithms for sensor network node localization with uncertainties

in anchor positions and/or propagation speed, IEEE Trans. Signal

Process. 57 (February (2)) (2009) 752763.
[20] L. Rui, K.C. Ho, Elliptic localization: performance study and optimum
receiver placement, IEEE Trans. Signal Process. 62 (September (18))
(2014) 46734688.