MAXIMUM LIKELIHOOD AND CRAMER-RA0 LOWER BOUND

ESTIMATORS for (NONLINEAR) BEARING ONLY PASSIVE

TARGET TRACKING

S.KOTESWARAM O , Member IEEE

Scientist IF', Naval Science and Technological lab, Visakhapatn,am-27, India

Abstract: Cramer- Rao lower bound (CRLB) is very recent contribution - a hybrid coordinate
a powerful tool in assessing the performance of system approach by Walter Grossman [4] are the
any estimation algorithm. S. C. Nardone at.al.,[l] successful contributions in this field. These
developed Maximum Likelihood EstimatorWLE) researchers tried to use (extended Kalman filter
for passive target tracking using batch (because of its advantages;) and at the same time
processing. In this paper, this batch processing is to eliminate filter divergence. Maximum
converted into sequential processing so that it is Likelihood is found to be a suitable algorithm for
useful for the above real time application using passive target tracking applications, by virtue of
bearings only measurements. Adaptively, the ils characteristics. S . C . Nardone at.al.,[ 11
weightage of each measurement is computed in developed Maximum Likelihood Estimator
terms of its variance and is used along with the W E ) for passive target tracking using batch
measurement, making the estimate a generalized processing. The solution of the gradient equation
one. Instead of assuming some arbitrary values, was obtained by Gauss-Newton iteration scheme.
Pseudo Linear Estimator outputs are used for the Cramer- Rao lower bound (CRLB) is a
initialization of MLE. The algorithm is tested in powerful tool in assessing the performance of the
Monte Carlo simulation and its results are filter. Let P be the estimation error covariance
compared with that of CRLB estimator. From the matrix corresponding to any unbiased estimator of
results, it is observed that this algorithm is also the unknown deterministic variables, then the
an effective approachfor the bearing only passive inequality can be stated as
target tracking. P2P =J
* -1
(1)
where J is the Fisher Information Matrix and P*
1. Introduction: defines the best that can be done in the above
sense [7]. In the present paper, MLE in batch
In the ocean environment, an observer monitors processing is converted in1o sequential processing
noisy sonar bearings from a radiating target, to suit underwater applications like real time
which is assumed to be traveling at a constant passive target tracking. All the elements of
velocity. The observer processes these covariance matrix are replresented recursively in
measurements and finds out range, course, terms of measurement equation. These terms are
bearing and speed of the target. Here the known as Recursive Sums and are maintained
measurement is nonlinear, making the whole throughout the algorithm. This approach avoids
process nonlinear. There are many methods the computational complexity by computing only
available [1]-[6] to o:tain target motion the incremental values for every new bearing
parameters in the above situation. The modified measurement. These incremental values are used
gain extended Kalman filter (MGEKF) developed to update the Recursive Sums in Covariance
by Song and Speyer [2], the modified polar matrix. Twenty Recursive Sums are to be updated
extended Kalinan filter by Aidala [3] and the on the arrival of new bearing measurement. This

0-7803-5148-7/98/$10.0001998 IEEE 44 1
method does not increase computation burden A -1 A
with any additional number of samples. R =diag[r(k)] (7)
CRLB is obtained by doing partial A A A A
derivative of the noise free measurement equation t cos B(1) - t sin B(1) cos B(l) - sin B(l)
w.r.to state [7], evaluated at the correct state. A A A A

With this simple modification, MLE structure is B(2) - 2t sin B(2) cos B(2) - sin B(2)
converted into CRLB structure. Adaptively, the A A A A
weightage of each measurement is computed in kt cos B(k) - ktsin B(k) cos B(k) - sin B(k)
terms of its variance and is used along with the
measurement, making the estimate a generalized (8)
where Z and h(Xs) are the vectors of measured
estimate. In underwater, the bearing measurement
and actual bearings and W is the input covariance
is corrupted with large amount of noise. Hence the
diagonal matrix. The solution to this nonlinear
measurements are averaged over an interval to
system of equations yields the Maximum
reduce the large variance of noise in the
Likelihood Estimate.
measurements. Instead of assuming some
arbitrary values, Pseudo L i x a r Estimator (PLE)
outputs are used for the initialization of MLE and
2.2 Maximum Likelihood Estimator in
CRLB. sequential processing: It is assumed that some
The algorithm is tested against several initial estimate of target state vector components
geometries in Monte Carlo simulation. For the is available. Let the solution be required at time
purpose of illustration, the results of one typical instant (k+l). The estimated range x and y
scenario are presented and compared with that of components are translated to time instant (k+l) as
CRLB. follows. Using the initial estimate of target state
vector, the following are rewritten
A A A
2.Mathematical Modeling: rx (1) = x(0) + t x- x* (1) (9)
A A A
The target state equation is given by
ryw = Y(O)+tY-Yo(l) 10)
x,(k + I ) = #o( + W X , (k) (2)
where Xs (k) is a state vector with target velocity A A A A
and position components and is given by where x,y, x(0) and y(0) are initial estimate of
target state vector components and xo and yo are
Y ( k ) x(k) YW observer position components. Let
A A
rxl (k + 1)and r (k + 1)represent
Yl
where 4( k + 11k)is a transient matrix . The
A A
bearing measurement is given by rx (1) and ry (1) at time instant (k+l). These are
B(k) = Tan" (rx (k)/ry (k)) (1) given by
A A A A
where rx(k) and ry(k) are the relative range
r x l ( k +1) = x(k +1)- (1+k)t X+ t x- xo(l)
components at instant k. The measured bearing, A A
Bm(k) is given by = x(k + 1) - kt X- x0 (1) (1 1)
, oc) = BQ + Y(k)
B (5) A A A

where y( k) is a zero mean Gaussian noise. r y l ( k +1) = y(k +1) -kt y- yo (1) (12)
In the above fashion the range components at all
2.1. Maximum Likelihood Estimator: time instants are translated to time (k+l). The
The gradient equation is given by [11 bearing residue at instant (k+l) is calculated as
follows. Let at instant (k+l), the observer is at
T A A-' A
AZ (Z,k)R W-'(Z-Z(XS)) = 0 point 0 and target is at point T ,as shown in Fig.
(6)
1. As actual bearing and range are not known,
where
estimated bearing B(k+1) and range ?(k +1)

442
are assumed to be actual bearing and range of the SSUMS[4]=
target.
G-F=E-
Bm(k + 1) - B(k + 1) =
n@) In the above fashion all the terms are rewritten in
iY (k + l)sinB, + 1) -ix(k + l)cosB, (k
(IC + 1) the form of SUMS. Initially all SUMS are zero
(13) and these are updated whenever the bearing
*+I) measurement is available. The estimated target
state vector at any instant is used to find out target
Using equation (13) the normal equations are motion parameters and range components
derived. For the purpose of presentation the first which
normal equation is given by

J
are required to update SUMS to estimate the
improved solution at next instant.

3.Gramer-Rao Lower Bound

Estimation :-
i (k+ lX-kt){xg(l)cosB,(l)- yg(l)sinB,(l)} Cramer-Rao Lower Bound (CRLB) is an ideal
- Y1
estimator. In this ideal case, the information
0 2(1)
matrix, C, becomes a noise free measurement
(14) matrix. The difference between realizations of
The normal equations, after getting k CRLB and MLE is in the construction of matrix
measurements are transformed into matrix form, ‘C’. In the case of CRLB, the SUMS
C Xs= D ,where corresponding to matrix ‘(2’ are developed with
$,I (k+1X-kt)2cosB,(1) i y 2 ( k + l X 1 - k ) 2 t 2cwB,(2) true measurements.
C(1,1)= T

0 2(1) Q *(2)
2 2
iy3(k+lX2-k) t cosB,(3) 4. Simulation results:
+ + . (1 5 )
Q 2(3)
(The other components are not shown to reduce Computer algorithm is constructed and tested
the no.of pages.) The above equations are not with simulated data to illustrate the performance
flexible for implementation in sequential of this multistage estimalor. All raw bearings
avialable at two seconds period are corrupted by
processing and so, are converted into sequential
additive zero mean Gaussian noise with a r.m.s
processing as follows. The procedure is shown for
levei of 1 degree and thein preprocessed over a
the simple term C(4,4). After the arrival of first
bearing measurement, C(4,4) can be written as period of twenty seconds Corresponding to a
+l)sinBm(l) scenario in which the target is at the initial range
of 19000 yards at initial bearing of zero degrees,
0 2(1) the target is assumed to be: moving at a constant
course of 140 degrees at a speed of 25 knots.
= I;(k+l)SSUMS[Oj -z(lc+l)ssuMs[lj -SSUMS[4j} Observer is assumed to be doing ‘S‘ maneuver on
(16) line of sight at a constant speed of 7 knots.
where Pseudolinear estimator generates initial estimate
for MLE. Finally the results have been ensemble
SSuMS[O]= +previous SSUMS[O] (17) averaged over several Monte Carlo runs. The
0 2(k>
results of MLE and CUE) are shown in Fig. 2
-t sinBm (k) with continuous lines and dotted lines
SSUMS[l]= +previous SSUMS[11
)2’ respectively. From the results we can observe that
(18)

443
the estimates of MLE approach to that of CRLB at target. In this paper the algorithm developed by
time index twenty fifth minute. Nardone et.al., [l] is extended with sequential
processing. The results of this MLE algorithm are
compared with that of the CRLB algorithm and
5. Summary and conclusion: from the results it is observed that MLE is also an
effective approach for bearings only passive target
MLE is proposed for developing bearings only tracking-
passive target tracking filter for an underwater

REFERENCES:

[11. S.C Nardone, A.G. Lindgren and K.F.Gong," Fundamental properties and performance of conventional bearings
only target motion analysis 'I, IEEE Trans. Automatic Control Vol. Ac -29, No.9, September 1984, PP 775-787.
[2] . T.L.Song & J.L. Speyer," A stochasticAnalysis of a modified gain extended kalman filter with applications to
estimation with bearing only measurements", IEEE Trans. Automatic Control Vol. Ac -30, N0.10, October 1985, PP
940-949.
[3] . V.J. Aidala and S.E.Hamme1, Utilization of Modified Polar Coordinates for bearings only tracking",IEEE
I'

Trans. Automatic Contrd Vol. Ac-28, N0.3, Mar& 1983, PP 283 - 294.
[4]. W. Grossman, Bearings only tracking : A hybrid coordinate system Approach", J.Guidance ,Vol. 17, No.3,
It

May-June,l994, PP 451457.
[5]. A.G.Lindgren, K.F.Gong," Position and Velocity Estimation via Bearing Observations", IEEE Trans. Aerosp.
ElectronSyst., vol. AES-14, July 1978, PP 564-577.
[6].A.G.Lindgren, K.F. Gong," Properties of a bearings only motion analysis estimator:An interesting case study in
system observability in Proc. 12th Asilomar Cod. Circuits, Syst., and comput., Novl978, pp50-58.
'I,

[7]. J.H. Taylor "Cramer-Rao Estimation Error Lower Bound Computation for Deterministic Nonlinear Systems",
IEEE Trans. Automatic control, vo1.A~-24,N0.2, April 1979, PP 343-344.

9 13 '17 2'1 25 2
9
Emin rrinutes

Fig. 1. Target Motion Analysis problem Fig.2(a)

10 10
Sera kra
mbec
5 - 5

0 0
9 13 17 21 25 29 9 13 17 21 25 29
l i m e in rrirutes 'line in " d e s

Fig.2(b) Fig.2(c)

444