Progress In Electromagnetics Research B, Vol.

30, 103116, 2011

H-INFINITY FILTER BASED PARTICLE FILTER FOR
MANEUVERING TARGET TRACKING
Q. C. Wang, J. Li, M. X. Zhang, and C. H. Yang
Department of Computer Science
Xiamen University, Xiamen, Fujian, China
AbstractIn this paper, we propose a novel H-innity lter based
particle lter (HPF), which incorporates the H-innity lter (HF)
algorithm into the particle lter (PF). The basic idea of the
HPF is that new particles are sampled by the HF algorithm.
Since the HF algorithm can fully take into account the current
measurements, when the new algorithm calculates the proposed
probability density distribution, the sampling particles can take
advantage of the system current measurements to predict the system
state. The particles distribution we obtained approaches nearer to the
state posterior probability distribution and the HPF alleviates the
sample degeneracy problem which is common in the PF, especially
when the maneuvers of the target tracking are large. Furthermore,
the HF algorithm can adjust gain imbalance factor by adjusting
disturbance attenuation factor, from that the new algorithm can
get the compromise between the accuracy and robustness and we
can obtain satised accuracy and robustness. Some simulations and
experimental results show that the proposed particle lter performed
better than the PF and the Kalman particle lter (KPF) in tracking
maneuvering target.
1. INTRODUCTION
For linear or Gaussian problems of the tracking algorithm, the Kalman
lter (KF) is widely used to get optimal solutions, and it can achieve
good tracking performance [1, 2]. Unfortunately, many practical
maneuvering target tracking problems are nonlinear or non-Gaussian.
In this case, a variety of tracking algorithms have been proposed to
Received 15 March 2011, Accepted 4 May 2011, Scheduled 6 May 2011
Corresponding author: Qicong Wang (qcwang@[Link]).
104 Wang et al.
evaluate the probability distribution, such as Extended Kalman lter
(EKF) and Unscented Kalman lter (UKF) [35]. Since the PF can
approach the Bayesian optimal estimate with innite samples, it is
more accurate than the EKF and UKF and is often chosen over the
EKF or UKF. In recent years, the PF has attracted many researchers
attention, which are also known as sequential Monte Carlo (SMC)
methods and can eectively deal with non-linear and non-Gaussian
problems [612]. Meanwhile, a large number of tracking methods about
the PF have been presented [1315]. Although the PF has been proven
successful in dealing with tracking maneuvering target, it has some
disadvantages, and one of which is that employing uniform re-sampling
leads to the particle impoverishment problem [1618]. To deal with
this problem in the lter, two key factors can be adopted, which are
the proper selection of the proposed distribution and the re-sampling
method mentioned in [19, 20]. In the light of the selection of proposed
distribution, the PF algorithm uses system state transition probability
as its importance density function. Since the density function, which
the PF adopted, does not utilize the latest measurements to generate
new particles, the result is that the produced particle samples focus
on the last of the posterior probability distribution, which leads to a
blind choice in the particles and makes the reduction of the ltering
precision. So the PF lter usually has unsatisfactory performance,
and sometimes the PF algorithm cannot be eectively utilized. For
this problem, people try to nd some other methods, which use latest
measurements to enhance performance, to generate new particles, such
as the KF, EKF and UKF [2123]. For the KF, EKF and UKF, they
cannot always get both high accuracy and robustness at the same
time, but the HF can get the compromise between the accuracy and
robustness by adjusting disturbance attenuation factor. In this paper,
we incorporate the HF algorithm [2428], which can fully take into
account the current measurements, into the PF for maneuvering target
tracking. The proposed HF based particle lter (HPF) both has
the inherent advantages of the PF and the HF and shows a marked
improvement in the maneuvering target tracking.
The layout of this paper is as follows: In Section 2, the PF
algorithm and Halgorithm are formulated; meanwhile, the proposed
HPF algorithm is presented in detail. Simulation results and
discussions are given in Section 3, and we conclude this paper in
Section 4.
Progress In Electromagnetics Research B, Vol. 30, 2011 105
2. H FILTER BASED PARTICLE FILTER
2.1. Basic Theory of Particle Filter
Because the particle lter has good performance in tracking nonlinear
and non-Gaussian problems, we employ it to solve the state estimation
problem. The particle lter rst starts with a number of particles
which are initialized. After that, each particle is generated by the
density function [29, 30].
Give a state space model:
x
k
= f(x
k1
, u
k1
) +w
k
(1)
z
k
= h(x
k
) +v
k
(2)
where, f(.) is the system dynamic function, h(.) is the system
observation function, x
k
and z
k
are the target state vector and target
observation vector at time k, respectively, w
k
is the Gaussian system
noise vector, v
k
is the Gaussian observation noise vector, we take k as
the time index. The detailed particle lter algorithm [5] is described
as follows:
step 1: Initialization. Sample the initial particles x
(1)
0
, x
(2)
0
,
. . . , x
(N)
0
, N is the number of particles and
i
0
= 1/N, i = 1, 2, . . . N.
step 2: Importance Computation. We get the predicted
particles at time k by bringing the particles at time k 1 into (1).
The importance weight of each predicted particle is computed through
(3) when we obtain the target observation at time k, and then, we
normalize the particle weight by the Equation (4):

i
k
=
i
k1
p(z
k
|x
i
k
)p
_
x
i
k
|x
i
k1
_
q
_
x
i
k
|x
i
k1
, z
k
_
(3)

i
k
=
i
k
/
N

j=1

i
k
(4)
where, q(x
i
k
|x
i
k1
, z
k
) is the importance density function, generally,
q(x
i
k
|x
i
k1
, z
k
) = p(x
i
k
|x
i
k1
). After normalizing the weight, we can
get the approximate posterior distribution p(x
k
|z
1:k
) by (5), (.) is the
Dirac function.
p(x
k
|z
1:k
) =
N

j=1

i
k
(x
k
x
i
k
) (5)
step 3: Resample. Accept the particles that have higher
importance weights, meanwhile, their cumulative probabilities are
bigger than a given threshold. Eliminate those particles that have
106 Wang et al.
lower importance weights and their cumulative probabilities are smaller
than the given threshold. Reset each particle weight
i
k
= 1/N, i =
1, . . . , N.
step 4: Output calculation. The posterior probability
estimation of the state is obtained approximately through (6):
x
k

N

i=1

i
k
x
i
k
(6)
step 5: k = k + 1, and move to step 2.
In the particle lter, we can choose some dierent proposed
distribution functions, and the most commonly used way is choosing
the prior density as its proposed distribution function [11, 12] in the PF,
that is, q(x
i
k
|x
i
k1
, z
k
) = p(x
i
k
|x
i
k1
). However, the shortcoming of this
method is that it does not consider the system current measurements,
which brings about the particle degeneracy problem. For now, two key
factors in preventing particle degeneracy [10] are the proper selection
of the proposed distribution and the re-sampling method. In the light
of the selection of proposed distribution, we need an algorithm that
can consider the system current measures. So we take advantage of
the H Filter and incorporate it into the particle lter.
2.2. Basic theory of H Filter
Generally, we can consider a time-varying discrete state model in the
Kerin space as following:
_

_
X(k) = (k)X(k 1) + (k)W(K)
Z(k) = H(k)X(k) +V (k)

S(k) = L(k)X(k), k = 1, 2, 3, . . .
(7)
where, X(k) R
n
is the system state vector at time k, Z(k) R
n
is
the observation vector at time k and

S(k) is the given state variable.
The matrixes (k), (k), H(k) and L(k) are preset to known matrixes
in general. We suppose the system noise W(k) and the observation
noise V (k) are energy bounded l
2
signals, that is,

k=0
||W(k)||
2
<
and

k=0
||V (k)||
2
< , where the sign ||.||
2
denotes the l
2
norm. We
have no hypothesis on their statistical properties. X(0), W(k) and
V (k) meet the following requirements:
_ _
X(0)
W(j)
V (j)
_
,
_
X(0)
W(k)
V (k)
_ _
=
_

0
0 0
0 I/
jk
0
0 0 R

(k)
jk
_
(8)
Progress In Electromagnetics Research B, Vol. 30, 2011 107
where, R

(k) =
_
I 0
0
2
f
I
_

jk
. For the given observations Z(k),
we denote

X(k) as the estimate of X(k), so we can obtain the state
estimation error e(k) :
e(k) =

X(k) X(k) (9)
The H ltering algorithm is a Kalman lter in Krein space actually,
so we design H lter based on (6) and (7) as follows:
X(k|k 1) = (k|k 1)X

(k 1) (10)
P(k|k 1) = (k|k 1)P(k 1)
T
(k 1)+(k1)
T
(k1) (11)
X

(k) = X(k|k1) +K

(k1)
_
Z(k)H(k)X(k|k1)

S(k)L(k)X(k|k1)
_
(12)
K

(k) = P(k|k 1)
_
H
T
(k) L
T
(k)

R
1
e
(k) (13)
P(k) = (I K

(k) [ H(k) L(k) ]) P(k|k 1) (14)

R
e
(k) =
_
I 0
0
2
I
_
+
_
H(k)
L(k)
_
P(k|k 1)
_
H
T
(k) L
T
(k)

(15)
The above formulas constitute the H robust ltering algorithm.
Because of using the dierent ltering gain algorithm, the H lter
is dierent from the standard kalman lter essentially. When the
disturbance attenuation factor , we can see that M(k) 0
through (16), which makes the H lter recursion degenerate into
the Kalman lter recursion. So the H norm of the Kalman lter
may be very large, which result in poor robustness performance. As
the disturbance attenuation factor min, we nd M(k) I
through (16), where I is the identity matrix, though we can get good
robustness, the estimation square error is quite large. So we can obtain
satisfactory requirement by adjusting the parameter according to
testing experiment in practice.
Under the Krein space conditions, suppose that M(k) is the
ltering gain imbalance factor, which describes the process that the
disturbance attenuation factor adjusts ltering gain. They meet the
following equation:
M(k) = (k)(
2
I + (k))
1
(16)
where, (k) = (I K
s
(k))H(k)P(k|k 1). From the analysis of the
algorithms and the above formulas, we employ the decomposition of
matrix and inverse theory in R
e
(k), and bring the results into the
formula (13), assuming L(k) = I, we get:
K

(k) = [ (I M(K))K
s
(k) M(k) ] (17)
108 Wang et al.
K
s
(k) = P(k|k 1)H
T
(k)(I +H(k)P(k|k 1)H
T
(k))
1
(18)
We can know by the above knowledge: the process that disturbance
attenuation factor adjusts the robustness of the lter can be described
as to adjust the gain of the lter process essentially.
K
s
(k)
M(k)
K

(k)
(14)
P(k)
(11)
P(k + 1|k)
(18)
K
s
(k + 1) (19)
2.3. Proposed HPF
The process of the HPF is shown in Table 1, where N is the total
number of particles. x
j
k
(k > 0) is the new particle generated by
HF. H prediction(.) is the function that performs the H lter
algorithm using the formulas (10)(15) described in Section 2.2. Pr(.)
is the probability that determines the resampling particle x
j
k
according
to the discrete weight distribution
j
k
.
3. SIMULATION RESULTS
3.1. Case 1: One-dimension Target Tracking With Large
Maneuvers
3.1.1. Target Scenario
For one-dimensional tracking problem, we adopt the following system
state space model, and at the same time we compare the performance
of the PF, KPF and HPF in terms of tracking accuracy.
_
x
t
= 0.5x
t1
+ 25x
t1
/
_
1 +x
2
t1
_
+ 8 cos[1.2(t 1)] +w
t
y
t
= x
2
t
/20 +v
t
(20)
where, w
t
and v
t
are the vector input white noise with zero mean,
x
0
= 0.1, and the particle number N is 50 and = 0.5. We have
implemented the algorithms in MatlabR2009a.
3.1.2. Tracking Performance Comparison
Figure 1 shows the tracking results by the PF, KPF and the proposed
HPF. We can see that the Kalman particle lter, and the proposed
method can well estimate the motion state of target throughout the
entire movement process.
Figure 2 shows the position error of estimated position
corresponding to PF, KPF and HPF. It is obvious that when the
target is during the maneuvering, the proposed HPF guarantees the
tracking accuracy and performed better than the KPF, but the PF
Progress In Electromagnetics Research B, Vol. 30, 2011 109
Table 1. The simple description of the HPF algorithm.
Initialization: k = 0
for j = 1, . . . , N
Sample x
j
0
from p(x
0
); Calculate the weight
j
0
= 1/N
end
for k = 1, 2, . . . (main loop)
H Prediction
for j = 1, . . . , N
x
j
k
=H prediction ( x
j
k1
)
end
Importance Step
for j = 1 : N, calculate the importance weights:

j
k
=
j
k1
p(z
k
|x
j
k
)p(x
j
k
|x
j
k1
)
q(x
j
k
|x
j
k1
, z
k
)
,
where, q(x
j
k
|x
j
k1
, z
k
) = p(x
j
k
|x
j
k1
).
end
Followed by normalization:
j
k
=
j
k
/

N
j=1

j
k
Resampling Step
for j = 1, . . . , N
Pr (x
i
k
= x
j
k
)=
j
k
,
Reset the weights
i
k
= 1/N, (i = 1, . . . N).
end
State Estimation Step
Calculate the desired estimate x
k
: x
k
=

N
j=1

j
k
x
j
k
end
cannot ensure good tracking accuracy. The reason is that the PF does
not think over the current measurements but the KPF and the HPF
employ the current measurements. To the KPF and the HPF,
HPF can get the compromise between the accuracy and robustness
by adjusting disturbance attenuation factor, so, the proposed algorithm
is the best among the three. From Fig. 3, we can see that the proposed
algorithm is much more robust than the other two and that the HPF
keeps the satisfactory results throughout the whole process.
110 Wang et al.
Figure 1. Estimated trajectory
by PF, KPF and HPF.
Figure 2. Position Error by PF,
KPF and HPF.
Figure 3. Position RMSE by PF, KPF and HPF with dierent
particle number.
3.2. Case 2: Two-dimension Target Tracking With Large
Maneuvers
3.2.1. Target Scenario
For two-dimensional tracking problem, we consider a relatively
complicated scenario. There are seven target motion patterns as shown
in Table 2 for our experiment in which we track a target with large
maneuvers and the period is longer. We set the initial motion pattern
as (u
x
= 28 m/s, u
y
= 28 m/s, = 0

/s). The initial position

is (1000 m, 1000 m) and = 100. The target motion is constant
velocity or constant turn in dierent time interval. Here, the particle
is described by a vector containing the position on X, the velocity on
X, the position on Y, the velocity on Y. The target is modeled by the
following model:
x
t
= F(, T)x
t1
+B(T)w
t
(21)
Progress In Electromagnetics Research B, Vol. 30, 2011 111
where, is the turn rate,
F(, T) =
_

_
1 sin(T)/ 0 (1 cos(T))/
0 cos(T) 0 sin(T)
0 (1 cos(T))/ 1 sin(T)/
0 sin(T) 0 cos(T)
_

_
is the state
transition matrix, B =
_

_
T
2
/2 0
T 0
0 T
2
/2
0 T
_

_
is the input matrix, w
t
is the
matrix of input white noise with zero mean.
Table 2 lists the detailed description of the target motion. The
target starts a constant velocity motion from position (1000 m, 1000 m)
with initial speed (u
x
= 28 m/s, u
y
= 28 m/s, = 0

/s) and = 100.

3.2.2. Tracking Performance Comparison
Figure 4 shows the tracking results by the PF, KPF and the proposed
HPF. Fig. 5 and Fig. 6 show the tracking results by the PF, KPF
and the proposed HPF on X and Y, respectively. It is clear that the
proposed method can well estimate the motion state of target.
Figure 7 shows the position error of estimated position
corresponding to the PF, KPF and HPF. We can see that though
the maneuvers are large, good performance can be obtained by HPF.
Figure 8 shows the position RMSE by the three lters with
dierent particle number. Form this gure, it is obvious that the
proposed algorithm is more precise and robust than the PF and KPF
even when the particle number is small.
Table 2. The process of target motion.
Time interval The target motion patterns
020 u
x
= 28 m/s, u
y
= 28 m/s, = 0

/s
21100 u
x
= 28 m/s, u
y
= 28 m/s, = 6

/s
101200 u
x
= 28 m/s, u
y
= 28m/s, = 0

/s
201240 u
x
= 28 m/s, u
y
= 28 m/s, = 9

/s
241300 u
x
= 28 m/s, u
y
= 28 m/s, = 0

/s
301400 u
x
= 28 m/s, u
y
= 28 m/s, = 7

/s
401440 u
x
= 28 m/s, u
y
= 28 m/s, = 0

/s
112 Wang et al.
Figure 4. Estimated trajectory
by the PF, KPF and HPF.
Figure 5. Estimated trajectory
on X by the PF, KPF and HPF.
Figure 6. Estimated trajectory
on Y by the PF, KPF and HPF.
Figure 7. Position error of
estimated position to the PF,
KPF and HPF.
Figure 8. Position RMSE by
the PF, KPF and HPF with
dierent particle number.
Figure 9. Position error of
estimated position to the KPF
and the HPF with dierent .
Progress In Electromagnetics Research B, Vol. 30, 2011 113
Figure 9 shows the position error of estimated position to the
KPF and the HPF with dierent values of , where the values of
are 10
2
, 10
3
, and 10
5
, respectively. When = 10
2
, we can get good
robustness and accuracy. But when = 10
3
or = 10
5
or larger, the
lter recursion may approach nearer to the KPF recursion which leads
to poor robustness properties.
4. CONCLUSION
In this paper, a novel HPF for maneuvering target tracking
has been proposed, and the simulation results demonstrate that
the new algorithm has better accuracy and robustness in tracking
maneuvering target. The proposed HPF algorithm incorporates
the HF algorithm into the standard particle lter, so the new
algorithm can fully take into account the current measures and make
the particles distribution more approach to the station posterior
distribution. Furthermore, the proposed algorithm can get the
compromise between the accuracy and robustness by adjusting
disturbance attenuation factor. Finally, simulation results demonstrate
that the proposed algorithm can achieve higher prediction precision
and better robustness, and meanwhile, the HPF is proved to be
eective and practicable in tracking the maneuvering target with large
maneuvers.
ACKNOWLEDGMENT
The work reported in this paper was supported by the National Natural
Science Foundation of China (No. 61001143).
REFERENCES
1. Bi, S. Z. and X. Y. Ren, Maneuvering target doppler-bearing
tracking with signal time delay using interacting multiple model
algorithms, Progress In Electromagnetics Research, Vol. 87, 15
41, 2008.
2. Turkmen, I. and K. Guney, Tabu search tracker with adaptive
neuro-fuzzy inference system for multiple target tracking,
Progress In Electromagnetics Research, Vol. 65, 169185, 2006.
3. Chen, J. M., X. Cao, Y. Xiao, and Y. Sun, Simulated annealing
for optimisation with wireless sensor and actuator networks,
Electronics Letters, Vol. 44, No. 20, 12081209, 2008.
114 Wang et al.
4. Ristic, B., A. Arulampalam, and N. Gordon, Beyond the Kalman
Filter-particle Filters for Tracking Applications, Artech House,
Boston, 2004.
5. Vander Merwe, R., A. Doucet, N. D. Freitas, and E. Wan,
The unscented particle lter, Advances in Neural Information
Processing Systems, Vol. 13, 2000.
6. Li, A., L. J. Zhong, and S. Q. Hu, Robust observation model
for visual tracking in particle lter, International Journal of
Electronics and Communications, Vol. 61, No. 3, 186194, 2007.
7. Gordon, N. J., A. Doucet, and N. D. Freitas, On sequential
monte carlo sampling methods for bayesian ltering, Statistics
and Computing, Vol. 10, 197208, 2000.
8. Li, Y., Y. J. Gu, Z. G. Shi, and K. S. Chen, Robust adaptive
beamforming based on particle lter with noise unknown,
Progress In Electromagnetics Research, Vol. 90, 151169, 2009.
9. Hong, S. H., Z. G. Shi, and K. S. Chen, Novel roughening
algorithm and hardware architecture for bearings-only tracking
using particle lter, Journal of Electromagnetic Waves and
Applications, Vol. 22, No. 23, 411422, 2008.
10. Chen, J. F., Z. G. Shi, S. H. Hong, and K. S. Chen, Grey
prediction based particle lter for maneuvering target tracking,
Progress In Electromagnetics Research, Vol. 93, 237254, 2009.
11. Zang, W., Z. G. Shi, S. C. Du, and K. S. Chen, Novel roughening
method for reentry vehicle tracking using particle lter, Journal
of Electromagnetic Waves and Applications, Vol. 21, No. 14, 1969
1981, 2007.
12. Shi, Z. G., S. H. Hong, and K. S. Chen, Tracking airborne targets
hidden in blind doppler using current statistical model particle
lter, Progress In Electromagnetics Research, Vol. 82, 227240,
2008.
13. Singh, A. K., P. Kumar, T. Chakravarty, G. Singh, and
S. Bhooshan, A novel digital beamformer with low angle reso-
lution for vehicle tracking radar, Progress In Electromagnetics
Research, Vol. 66, 229237, 2006.
14. Shi, Z. G., S. Qiao, K. S. Chen, W. Z. Cui, W. Ma, T. Jiang,
and L. X. Ran, Ambiguity functions of direct chaotic radar
employing microwave chaotic Colpitts oscillator, Progress In
Electromagnetics Research, Vol. 77, 114, 2007.
15. Bugallo, M. F., J. Miguez, and P. M. Djuric, Positioning by
cost reference particle lters: Study of various implementations,
The International Conference on Computer as a Tool, EUROCON
Progress In Electromagnetics Research B, Vol. 30, 2011 115
2005, 16101613, 2005.
16. Miodrag, B., S. J. Hong, and M. D. Petar, Finite precision eect
on performance and complexity of particle lters for bearing only
tracking, Proceedings of the Conference on Signals, Systems and
Computers, Asilomar, 838842, CA, 2002.
17. Blom, H. A. P. and E. A. Bloem, Particle ltering for
stochastic hybrid systems, Nation Aerospace Laboratory, 43rd
IEEE Conference on Decision and Control , 32213226, IEEE,
Netherlands, 2004.
18. Arulampalam, M. S., S. Maskell, N. Gordon, and T. Clapp,
A tutorial on particle lters for online nonlinear/non-Gaussian
Bayesian tracking, IEEE Transactions on Signal Processing,
Vol. 50, No. 2, 174188, 2002.
19. Liuand, J. S. and R. Chen, Sequential monte carlo methods
for dynamic systems, Journal of the American Statistical
Association, Vol. 443, No. 93, 10321044, 1998.
20. Zaugg, D. A., A. A. Samuel, D. E. Waagen, and H. A. Schmitt,
A combined particle/Kalman lter for improved tracking of beam
aspect targets, IEEE Workshop on Statistical Signal Processing,
535538, 2003.
21. Turkmen, I. and K. Guney, Tabu search tracker with adaptive
neuro-fuzzy inference system for multiple target tracking,
Progress In Electromagnetics Research, Vol. 65, 169185, 2006.
22. Xu, S. S., F. M. Bugallo, and M. D. Petar, Performance
comparison of EKF and particle ltering methods for maneuvering
targets, Digital Signal Process (2006), Vol. 10, No. 001, 113,
2006.
23. Wang, J. J. and Q. Chatym, Object tracking by multi-degrees of
freedom mean shift procedure combined with the Kalman particle
lter algorithm, Proceedings of the 2006 International Conference
on Machine Learning and Cybernatics, 37933797, Dalian, China,
2006.
24. Xie, L., L. Lu, D. Zhang, and H. Zhang, Improved robust H2 and
H ltering for uncertain discrete-time systems, Automatica,
Vol. 40, No. 5, 873880, 2004.
25. Babak, H., H optimality of the LMS algorithm, IEEE
Transactions on Signal Processing, Vol. 44, No. 2, 267280, 1996.
26. Jin, S. H., J. B. Park, and K. K. Kim, Krein space approach to
decentralised H state estimation, IEE Proceedings Control
Theory and Applications, 502508, IEE, 2001.
27. Rami, S. M., D. A. Brent, and C. George, Stochastic
116 Wang et al.
interpretation of H and robust estimation, Proceedings of the
33rd Conference on Decision and Control , 39433948, IEEE, 1994.
28. Xu, B., J. Wei, and P. Q. Yan, Application of robust H
ltering in TV tracking system for maneuvering targets, IEEE
International Conference on Control and Automation Guangzhou,
756760, 2007.
29. Djuric, P., J. Koteeha, J. Zhang, Y. Huang, T. Ghinnai,
M. Bugallo, and J. Miguez, Particle ltering, IEEE Signal
Processing Magazine, Vol. 20, No. 5, 1938, 2003.
30. Gustafsson, G., F. Gunnarsson, N. Bergman, U. Forssell,
J. jansson, R. Karlsson, and P. J. Nordlund, Particle lters
for positioning, navigation, and tracking, IEEE Transations on
Signal Processing, Vol. 50, No. 2, 425436, 2002.