Sie sind auf Seite 1von 3

IEICE TRANS. FUNDAMENTALS, VOL.E86A, NO.

12 DECEMBER 2003

3291

LETTER

Constrained Location Algorithm Using TDOA


Measurements
Hing Cheung SOa) and Shun Ping HUI , Nonmembers

SUMMARY One conventional technique for source localization is to utilize the time-dierence-of-arrival (TDOA) measurements of a signal received at spatially separated sensors. A simple
TDOA-based location algorithm that combines the advantages of
two ecient positioning methods is developed. It is demonstrated
that the proposed approach can give optimum performance in geolocation via satellites at dierent noise conditions.
key words: positioning algorithm, weighted least squares, timedierence-of-arrival

to form a new computationally ecient location algorithm. It is shown that the proposed method is superior to the LCLS and its performance can attain the
Cramer-Rao lower bound (CRLB), which gives a lower
bound on the variance attainable by any unbiased estimator.

1.

Let (xl , yl , zl ), l = 1, 2, , L, with L 5, and


(xs , ys , zs ), be the known position of the lth sensor
and the unknown source location, respectively, and denote the TDOA with respect to the rst sensor by Dl1 ,
l = 2, 3, , L. The corresponding range dierences
{rl1 } are given by

Introduction

Time-dierence-of-arrival (TDOA) estimation between


signals received at spatially separated sensors is a frequently addressed problem in the signal processing literature [1]. An important application of time delay
estimation is to locate a radiative source in the areas
of radar, sonar [2], global positioning system (GPS)
[3], and more recently, wireless communications [4] and
speech processing [5], [6]. The TDOA information denes a set of hyperbolic equations that are highly nonlinear, from which the source can be located with the
knowledge of the sensor positions. This is not an easy
task and previous work relies on an iterative Taylorseries method [7] which is computationally expensive
and requires an initial guess close to the true position, indicating the possibility of local convergence.
In order to avoid solving the hyperbolic equations directly, they can be reorganized into a set of spherical equations by introducing an intermediate variable,
which is a function of the source position [2], [6]. Chan
and Ho have proposed [2] to use two weighted least
squares (WLSs) to solve for the source position with
exploiting the known relation between the intermediate
variable and the position coordinates implicitly, while
[6] incorporates this relation explicitly and the solution is found via the technique of Lagrange multipliers. We refer to these two approaches as the quadraticcorrection least-squares (QCLS) and linear-correction
least-squares (LCLS), respectively. In this Letter, the
advantages of the QCLS and LCLS will be utilized
Manuscript received November 12, 2002.
Manuscript revised August 5, 2003.
Final manuscript received August 19, 2003.

The authors are with the Department of Computer


Engineering and Information Technology, City University
of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong,
China.
a) E-mail: ithcso@cityu.edu.hk

2.

Proposed Algorithm

rl1 = cDl1 ,

l = 2, 3, , L

(1)

where c is the speed of signal propagation. In practice, the range dierence measurements obtained are
corrupted by noise:
rl1 = rl1 + nl1

(2)

where nl1 , l = 2, 3, , L, are assumed to be zero-mean


noises. From (1), we have

rl1 + (xs x1 )2 + (ys y1 )2 + (zs z1 )2

= (xs xl )2 + (ys yl )2 + (zs zl )2 ,
l = 2, 3, , L
(3)
Nevertheless, the solution for the (L 1) hyperbolic
equations is very sensitive to measurement noise and involves huge computations. Following [2], [6], we square
both sides of (3) and introduce an intermediate variable, namely,

(4)
Rs = (xs x1 )2 + (ys y1 )2 + (zs z1 )2
to yield the following set of linear equations:
(xs x1 ) (xl x1 ) + (ys y1 ) (yl y1 )
+(zs z1 ) (zl z1 ) + rl1 Rs

1
2
(xl x1 )2 + (yl y1 )2 + (zl z1 )2 rl1
,
=
2
l = 2, 3, , L
(5)
Formulating the problem in matrix form we have
G = h

(6)

IEICE TRANS. FUNDAMENTALS, VOL.E86A, NO.12 DECEMBER 2003

3292

where

G=

x 2 x1
x3 x1
..
.
xL x 1

y2 y1
y3 y1
..
.

z2 z1
z3 z1
..
.

r21
r31
..
.

yL y1

zL z1

rL1

f () = pT ( + I)2 q

2
(xL x1 )2 + (yL y1 )2 + (zL z1 )2 rL1

and the parameter vector = [


xs x1 , ys y1 , zs
s ]T consists of the estimates of the source position
z1 , R
as well as Rs .
In the QCLS, a coarse estimate is rst estimated
via minimizing a WLS function:
(7)

The weighting matrix W is given by


W = B Q BT

pi qi

( + i )2
( + i )2 i=1

(12)

where I is the 4 4 identity matrix, p = [p1 , p2 ,


p3 , p4 ]T = UT GT W1 h and q = [q1 , q2 , q3 , q4 ]T =
U1 GT W1 h. Unlike [6] which uses the secant
method to solve for f () and a two-step procedure to
estimate , we propose the following steps for source
location:
1. Set W = I.
2. Find the root of (12) whose value is closest to zero
by using Netwons method with zero as the initial
using (11).
guess. Then determine

s in
3. Construct W using (8) using the obtained R
and repeat Step 2.
It is noteworthy that the estimation procedure in [6]
can be simplied as Steps 1 and 2 without any loss
in performance.

(8)

o
o
, r31
,
where B is a diagonal matrix of the form diag(r21
o
o
s , and Q is the noise co , rL1 ) with rl1 = rl1 + R
variance matrix which is assumed known a priori. A
second estimate of is then obtained via minimizing
s2 based on anys y1 )2 + (
zs z1 )2 R
(
xs x1 )2 + (
other WLS procedure. Note that a few iterations for
the two steps are required to attain the best solution
s is not available at the beginning. On the
because R
other hand, the LCLS aims to minimize the standard
least squares of (G h)T (G h) subject to the constraint of

s2
ys y1 )2 + (
zs z1 )2 = R
(
xs x1 )2 + (

4

i=1

(G h)T W1 (G h)

( + i )2

i=1

2
(x2 x1 ) + (y2 y1 ) + (z2 z1 )2 r21
2
2
2
2

1 (x3 x1 ) + (y3 y1 ) + (z3 z1 ) r31


h=
..
2
.
2

(9)

and the solution is solved by the method of Lagrange


multipliers. In short, the QCLS is more complex but
can achieve the CRLB while the LCLS is computationally simpler but may not be optimum.
Combining the merits of both methods, we propose
to minimize (7) subject to (9) using the technique of
Lagrange multipliers. The cost function is of the form:

3.

Results and Conclusions

Computer simulations had been conducted to evaluate the performance of the proposed location algorithm
by comparing with the LCLS and the CRLB [2],[6].
We considered a GPS scenario with (xs , ys , zs ) =
(3895818.244, 325592.485, 5022723.016) m, (x1 , y1 , z1 )
= (8787698.532, 18691292.353, 16789171.19) m, (x2 ,
y2 , z2 ) = (2228828.313, 18073221.126, 19405688.467)
m, (x3 , y3 , z3 ) = (9775123.349, 8627055.585,
23247777.886) m, (x4 , y4 , z4 ) = (23770926.545,
11819578.480, 2041166.259) m, (x5 , y5 , z5 ) =
(14528356.708, 6175808.909, 21148031.464) m, (x6 ,
y6 , z6 ) = (1069123.922, 16400053.194, 20698995.486)
m and (x7 , y7 , z7 ) = (22653816.587, 13572760.609,
1638758.941) m. All simulation results provided were
averages of 1000 independent runs.
Figure 1 shows the mean square range errors
(MSREs) of the proposed method and LCLS as well

L(, ) = (G h)T W1 (G h) + T
(10)
where is the Lagrange multiplier to be determined
and = diag(1, 1, 1, 1). Following [6], we dierentiate L(, ) with respect to and then equate the results
to zero, which gives the estimate of as

= GT W1 G + 1 GT W1 h
(11)

Diagonalizing GT W1 G as UU1 where =


diag(1 , 2 , 3 , 4 ), and with the use of (10) and (11),
can be solved from:

Fig. 1 Mean square range errors for correlated noises with


equal variances.

LETTER

3293

been developed. It is shown that the proposed method


can achieve optimum performance and outperforms the
LCLS.
Acknowledgement
The work described in this paper was supported by a
grant from the Research Grants Council of the Hong
Kong Special Administrative Region, China [Project
No. CityU 1119/01E].
References
Fig. 2 Mean square range errors for uncorrelated noises with
unequal variances.

as the CRLB versus average noise power when nl1 =


nl n1 where all nl , l = 1, 2, , L, were white Gaussian noises with identical variances. We observe that
the performance of the WLS estimator met the CRLB
and was superior to that of the LCLS by approximately 1 dB, which corresponded to 20% reduction in
the MSREs, for the whole range of noise power. The
previous test was repeated for uncorrelated Gaussian
noises with Q = diag(n2 1 , n2 1 , n2 1 , n2 2 , n2 2 , n2 2 ) where
n2 2 = 100n2 1 and the results are plotted in Figure 2.
Again, the proposed method was optimum, and signicantly outperformed the LCLS by approximately 8 dB.
To conclude, a TDOA-based location algorithm
via combining the merits of the QCLS and LCLS has

[1] G.C. Carter, Coherence and Time Delay Estimation: An


Applied Tutorial for Research, Development, Test, and
Evaluation Engineers, IEEE Press, 1993.
[2] Y.T. Chan and K.C. Ho, A simple and ecient estimator for hyperbolic location, IEEE Trans. Signal Process.,
vol.42, no.8, pp.19051915, 1994.
[3] K.C. Ho and Y.T. Chan, Solution and performance analysis of geolocation by TDOA, IEEE Trans. Aerosp. Electron. Syst., vol.29, no.4, pp.13111322, 1993.
[4] J.J. Caery, Wireless location in CDMA cellular radio systems, Kluwer Academic Publishers, 1999.
[5] S.L. Gay and J. Benesty, Acoustic Signal Processing for
Telecommunication, Kluwer Academic Publishers, 2000.
[6] Y. Huang, J. Benesty, G.W. Elko, and R.M. Merereau,
Real-time passive source localization: A practical linearcorrection least-squares approach, IEEE Trans. Speech
Audio Process., vol.9, no.8, pp.943956, 2001.
[7] D.J. Torrieri, Statistical theory of passive location systems, IEEE Trans. Aerosp. Electron. Syst., vol.20, no.2,
pp.183197, 1984.

Das könnte Ihnen auch gefallen