I. INTRODUCTION
The task of traditional target tracking is to establish target kinematic
trajectories from sequences of noisy kinematic measurements in the
presence of false alarms and countermeasures [1], [2]. Ground moving
target indicator (GMTI) radar has been a major sensor for air-to-ground
target tracking. However, difculty arises in traditional GMTI tracking
when target density becomes high and targets move together, which
could result in merged tracks and switched track identities (IDs). The
reason for this is that traditional tracking uses only the normalized distance between a measurement and the predicted target location to decide if the measurement was potentially originated from the target (this
is sometimes called data association), which could quickly lose its capability if targets stay together within the sensor uncertainty range for
an extended period of time. With the advance in sensor devices such
as synthetic aperture radar (SAR) and HRR [13], additional information regarding target identication becomes available, which could be
very valuable in helping data association. Feature aided tracking (FAT)
is a new research area and has been gaining a great deal of attention
in recent years due to its signicant advantage over traditional target
tracking, especially for stressing and complicated scenarios. Besides
the research done by the authors [3], [6], [7], references [5] and [12]
addressed mainly the aspect of feature extraction for tracking and [11]
dealt with mostly moving target classication. So far, the features used
in feature aided tracking are based purely on target spatial information
and no temporal correlation information is explored. However, since
features are extracted from a continuous string of signature signals from
the same targets moving over time, how a feature changes over time is a
feature itself which is called motion feature in this note. However, it
is very difcult, if not impossible, to characterize a motion feature with
respect to the free time variable in the kinematics space; some kind of
normalization approach is needed to bound the space of motion feature in order for it to be useful. To handle this difculty, we introduce
a concept of local motion and dene a local motion feature which is
well behaved in a normalized relative motion space. Instead of a time
Manuscript received October 1, 2003; revised February 24, 2004 and October
4, 2004. Recommended by Associate Editor J. Hespanha.
L. Hong and N. Cui are with the Department of Electrical Engineering, Wright State University, Dayton, OH 45435 USA (e-mail:
lhong@cs.wright.edu).
M. Pronobis and S. Scott are with the Information Directorate, Air Force Research Laboratory, Rome, NY 13441 USA.
Digital Object Identier 10.1109/TAC.2004.841119
128
Fig. 3. HRR prole is approximated by pseudo range energy bins and a center bin.
target identication. Both local motion features and GMTI measurements are fed into a probabilistic logic based tracker where the data
association weights are jointly determined by both kinematics and feature information. Signicant tracking performance improvement over
the traditional tracking for the stressing scenarios has been achieved.
This note serves as a rst step toward a new research paradigm of complete HRR/GMTI information exploitation for feature aided tracking.
II. HRR LOCAL MOTION FEATURE EXTRACTION
A. HRR Prole Formation
Fig. 1.
An HRR prole is an instant one-dimensional (1-D) target signature in the range domain, which describes the distribution of reected
signal energy at a particular viewing angle called the aspect angle (or
pose angle) and a xed depression angle, Fig. 1. Because of the well
studied phenomenon of electromagnetic wave reection, target structure information can be decoded from an HRR prole, which in turn
is connected to the target identication. Since an HRR is actually an
instant snapshot of a target, it is quite suitable to be used for moving
targets. Interested readers can refer to [13] for details on HRR.
B. Global and Local Motions of a Moving Rigid Target
Fig. 2.
129
the decomposition of the composite motion into local and global motions is not unique. Extracting local information from 1D HRR proles is even harder. The next section will develop a method of deriving
local motion information from 1-D HRR proles.
ranges and the local ranges can be derived by removing the global
range which is the center bin range from the composite ranges. The
local ranges are
Ll
L
0 Lc
(4)
Lc
lmax + lmin
(1)
where lmax is the farthest range bin containing signal energy and lmin
is the nearest range bin having signal energy. We now dene the pseudo
range energy bins for the prole taken at aspect angle and depression
angle
using the method of matching moment of inertia
N
L
i=1
N
li2 ai
j =1
(2)
aj
L1
i=1
N
j =1
li2 ai
and
aj
L2
L1l
0 Lc
L1
and L2l
0 Lc
L2
(5)
l
@L
@
and al =
l
@ 2 L
2
@
(6)
where li is the ith range bin of the HRR prole, ai is the signal magnitude at the li range bin, and N is the total number of range bins containing signal energy. Equation (2) calculates a single pseudo range for
a one energy bin approximation as shown in Fig. 3. For a two energy
bin approximation, similar calculations can be carried out over odd and
even bin numbers of the prole, which are
N
i=1
N
j =1
li2 ai
(3)
aj
A higher number energy bin approximation can also be dened similarly; but for feature aided tracking, one or two energy bin approximation is proven to be sufcient. The pseudo ranges are still composite
1We have proven that with the knowledge of correspondence, one can
completely reconstruct a three-dimensional (3-D) target structure and calculate
global and local motions from a sequence of 1-D HRR proles.
2Different denitions of HRR energy bins and center bin will end up with
different local motion features, but information contained in these local motion
features is essentially the same.
After extraction of local motion features from HRR proles, the remaining question is how to combine them with target kinematic behaviors for tracking. In our probabilistic logic based tracker, the combination is performed through their likelihood functions. This section
describes the likelihood calculation of local motion features.
Using the HRR training data set of N types of targets, the local motion feature trajectories over angles for N targets can be described by
f = [
T
; T ; . . . ; T
T1
; T2
(7)
; . . . ; Tn
(8)
and
Tj
((j )
(9)
130
P
Fig. 4. Filterbank for lowpass ltered, highpass ltered, and high-highpass
ltered local ranges.
lterbank,3
T
f z
= exp
00:5 R0
0
(10)
where = z 0 T and T is the local motion features vector for
type Ti from the training data at the same angle index as the feature
measurement and R is a weighting matrix either derived from the
training process or provided by the user.
t
Zg =
l
zg
m
l=1
and Zt = z l
t;i
z kjk01 =
^
(11)
l=1
Zg =
t=1
t
Zg
and Z =
t=1
t
Z
(12)
The total measurements at time k for all targets and the total cumulative
GMTI/HRR measurements up to time k are
M
Zk =
t=1
t
t
Zg Z
(13)
l
l
zg z
k =
l=1
t
l :
(14)
(15)
t 0
t
t t;i
St;i
k = Hk Pkjk0 Hk + Rk
and
(16)
(17)
where Fkt 01 is a CA kinematic matrix and Hkt is the Jacobian matrix for the nonlinear measurement model. Gating is performed for each track based on predicted measurements, z^t;i
kjk01 ,
t;i
and their error covariances, Sk .
2) Type probability update:
t
i = P
=
t k
Ti Z
c1
1
c1
t k
Ti Z
=P
t k01
p Z T i ; Z
P
t t
p Z T i i
(18)
t k01
Ti Z
(19)
t
p Z T i =
m
l=1
1
l
t
p z Ti =
c2
m
l=1
l
T
f z
(20)
i k
k Z
i
k
T ;T ;...;T
Given all GMTI and HRR measurements, the track of target t with
its type being Ti is calculated by
t;i
t;i k
x
^ kjk = E xk Z
Hkt x^t;i
kjk0
Fkt 0 x^t;i
and
k0 jk0
t
t;i
Fk0 Pk0 jk0 Fkt 0 0 + Qkt 0
2P
t
T
M
t=1
T ;T ;...;T
P
t=1
t
T
t
T
t=1
;Z
jZ k
i
k
t
k
T Z
t=1
;Z
(21)
131
Fig. 5.
(Left) Estimated tracks using local motion feature aided tracking. (Right) Estimated tracks using traditional JPDA tracking.
ik
P
=
c3
c3
ik ;
ik
!
c03 V
m
l=1
P j 3 I 0 W H P j 01
t;i
kk
M
t=1
t=1
M
t=1
jZ 0
M
t=1
; Z k01
; Z k01
N z lk : z^tkjk01 ; Skt
l=1
f z t jT
PDt
0 PDt 10
(22)
lt
P tl jZ k
P ik jZ k !^ lt
ik
(23)
t
i
where !
^ l (k ) is a matrix representation of a feasible event [1].
4) Track update: The track update is performed as
t;i
x^t;i
^
kjk = x
kjk01 +
and
Pj
t;i
kk
0t
t
k
m
l=1
lt lt
P j 01 0 0 P j 3
00
W
t;i
kk
t
k
l=1
(24)
t;i
kk
t
l
t
l
t
l
t
k
kt
t
k
t
k
t
k
t;i
kk
and kt
m
=
l=1
lt lt
(26)
k 1
t
k
where
tT
; Z k01
Z
2P
=
t=1
;Z
where
p Zg jik ; tT
2p
=
tT
(25)
4Actually, GMTI measurements and HRR prole features are lightly coupled
through a complicated nonlinear function.
132
TABLE I
RMS ERRORS OF LOCAL MOTION FEATURE AIDED TRACKING (LMFAT) AND JPDA (50 RUNS)
Fig. 6.
V. CONCLUSION
TABLE II
COMPUTATIONAL COMPLEXITY COMPARISON OF LMFAT AND JPDA
feature aided tracking using one and two energy bin approximations
are shown. It can be seen that using the one range bin approximation
for HRR proles is sufcient enough for local motion feature aided
tracking. For comparison, Fig. 5 (right) shows the results of a traditional joint probabilistic data association (JPDA) tracker with using
kinematics information only. As expected, the tracks get switched at
the crossing point and the estimated track quality is degraded during
the parallel motion period. The RMS errors of the JPDA tracker are
also given in Table I for comparison.
To show how target ID information helps data association, we plot
target ID probabilities for tracks 13 in Fig. 6. It can be seen that the
target ID probabilities converge to around 0.9 quickly from the initial
number of 0.5, which explains why consistent ID information from
HRR proles could help data association in kinematic measurements.
It is expected that the performance improvement of local motion
feature aided tracker is achieved at a cost. Table II presents the CPU
times of each algorithm for one run (200 scans). It can be seen that the
LMFAT consumes about 80% more CPU time than the JPDA does.
Local motion features of a rigid moving target contain rich ID information and are very suitable for feature aided tracking. In this note, a
novel algorithm of local motion feature aided tracking is developed to
effectively combine information from both GMTI and HRR. The local
motion features of HRR proles capture essential structural information of specic targets over angles. These features are extracted by a
wavelet lterbank over a sliding window of length three. By combining
the local motion feature information with kinematics information, a
probabilistic logic based tracker is developed, which successfully separates the different tracks in a confusing scenario, as illustrated in the
simulations.
REFERENCES
[1] Y. Bar-Shalom and T. E. Fortmann, Tracking and Data Association. San Diego, CA: Academic, 1988.
[2] S. S. Blackman, Multiple Target Tracking With Radar Applications. Norwood, MA: Artech House, 1986.
[3] E. Blasch and L. Hong, Simultaneous feature-based identication and
track fusion, in Proc. IEEE Int. Conf. Decision and Control, Tampa,
FL, Dec. 1998, pp. 239244.
[4] I. Daubechies, Orthonormal bases of compact supported wavelets,
Commun. Pure Appl. Math., vol. XLI, pp. 909996, 1988.
[5] W. R. Franklin and R. R. Kalman, Complex spatial lters for automatic
target recognition and feature aided tracking, in Proc. SPIE Algorithms
for Synthetic Radar Imagery, vol. 4053, 2000, pp. 501512.
133
[6] B. Gu and L. Hong, Tracking 2-d rigid targets with invariant constraints, Inform. Sci., vol. 138, pp. 7997, 2001.
[7] L. Hong, S. Cong, M. T. Pronobis, and S. Scott, Wavelets feature aided
tracking (WFAT) using GMTI/HRR data, Signal Process., vol. 86, pp.
26832690, 2003.
[8] L. Hong, Multiresolutional distributed ltering, IEEE Trans. Autom.
Control, vol. 39, no. 4, pp. 853856, Apr. 1994.
, Multirate interacting multiple model ltering for target tracking
[9]
using multirate models, IEEE Trans. Autom. Control, vol. 44, no. 7, pp.
13261340, Jul. 1999.
[10] L. Hong and N. Cui, An interacting multipattern probabilistic data association algorithm for target tracking, IEEE Trans. Autom. Control,
vol. 46, no. 8, pp. 12231236, Aug. 2001.
[11] D. H. Nguyen, J. H. Kay, B. J. Orchard, and R. H. Whiting, Classication and tracking of moving ground vehicles, Lincoln Lab. J., vol. 13,
no. 2, pp. 275308, 2002.
[12] J. Schmitz and J. Greenewald, Model-based range extent for feature
aided tracking, in Proc. IEEE Int. Radar Conf., Alexandria, VA, May
2000, pp. 166171.
[13] D. R. Wehner, High-Resolution Radar, 2nd ed. Norwood, MA: Artech
House, 1995.