Prepared for the Second International Conference on

Information Fusion: FUSION99

Sunnyvale, CA, July 1999
1
Track Association and Track Fusion with Non-Deterministic Target Dynamics
Shozo Mori, William H. Barker
smori@raytheon.com, bbarker@raytheon.com
Raytheon Systems Company,
Advanced C
3
I Systems
San Jos, CA
Chee-Yee Chong
chong_chee@bah.com
Booze-Allen & Hamilton, Inc.,
San Francisco, CA
Kuo-Chu Chang
kchang@gmu.edu
George Mason University,
Dept. of Systems Engineering,
Fairfax, VA
Abstract - This paper compares various track fusion
algorithms and track association metrics, using a simple
linear-Gaussian-Poisson model, to examine their
performance under various degrees of non-
deterministicitity of the target dynamics, i.e., process
noises. Track fusion algorithms are compared by an
analytical method while track association metrics are
evaluated by Monte Carlo simulations.
Keywords: Distributed Data Fusion, Distributed Tracking,
Track-to-Track Association, Track Fusion, Non-
Deterministic Dynamics
1. Introduction
In many modern information gathering systems with
multiple physically distributed sensors, a distributed
data fusion architecture possesses several advantages
over a centralized architecture. Among them is
avoidance of data flow/processing bottle necks
through well-designed data distribution and
processing, without any significant loss of optimality
achieved by a centralized architecture with infinite
processing power. In the last three decades,
distributed information processing algorithms have
been one of the most studied areas in data processing.
In the field of target tracking, according to [1], we
can trace a pioneering work in distributed processing
back to Singers 1971 paper [2].
A general theory of distributed tracking was
built ([4] [7]) based on a general theory of
distributed estimation described in [3] and on a
general theory of multi-target tracking described in
[9] (which generalizes the multi-hypothesis tracking
algorithm developed by D. B. Reid [8]). Recently
this general distributed tracking theory was also
described in the random-set formalism ([10],[15])
using the general theory of multi-target tracking re-
written in such formalism ([10] [14]). This theory
is general enough to be applicable to almost any kind
of information flow pattern
1
and to a very wide class
of target and sensor models. However, it was pointed
out from the very beginning of its development that
there is difficulty in applying this theory to non-
deterministic target models.

1
including dynamically changing information flows.
In a traditional sense of distributed tracking,
track-to-track association and track fusion
(distributed filtering) replaces measurement-to-track
association and dynamic state estimation (i.e.,
filtering) in a single-site or centralized tracking. In
the area of distributed filtering, there have been a
large volume of papers and reports that describe
various distributed filtering algorithms [16] [25].
Also, in this traditional framework, it was recognized
that effects of target models non-deterministicity
cause certain difficulties and several attempts have
been made to study such effects and develop
algorithms to alleviate such difficulties, as shown in
[26] [29]. In addition, efforts to expand the general
framework mentioned before to cope with non-
deterministic target dynamics were also made ([30]).
The objective of this paper is to explore
issues related to non-deterministic target dynamics in
search of effective distributed tracking algorithms.
This paper characterizes and evaluates several
representative algorithms brewed through almost
three decades of development in distributed tracking
as described above. For this purpose, we will take a
traditional approach to distributed tracking, i.e., we
will discuss track-to-track association (Section 3) and
track (state-estimate) fusion (merge) (Section 2)
separately. Furthermore, in order to make evaluation
and comparison easy, we will restrict ourselves to a
linear gaussian case with the simplest information
network, i.e., two sensors with local processors and
one central processor.
2. Track Fusion (Distributed Filtering)
Algorithms
There is a wide class of distributed filtering
problems, and even more ways to describe them. In
this paper, however, we will consider the simplest
case in which two tracks generated by two local data
processing nodes are to be fused together.
2.1. Problem Statement: Let us assume that the
state
t
x at time t of a target is described by a linear
stochastic differential equation,
t t t t t
dw B dt x A dx + on an interval ) , [
1
t . As
2
usual, the initial condition on
1
t
x is given as a
gaussian random vector independent of the unit-
intensity standard Wiener process ( )

1
t t t
w . We
assume that matrices
t
A and
t
B with compatible
dimensions satisfy all the necessary regularity
assumptions to guarantee unique existence of a
solution ( )

1
t t t
x to the linear stochastic equation.
Assume two sensors observe the target states as
ik
k
t ik ik
x H y + at given time
2
,
n
t t t < < < L
2 1
,
for each sensor i, with appropriate matrices
ik
H and
independent additive measurement noises
ik
, i.e.,
zero-mean gaussian vectors with positive definite
3
variance matrix
ik
R .
Define a local target state estimate as any
conditional mean ( )
,
_

k
j
j i
y t x
1
of target state
t
x at
time t given a cumulative local measurement
( )
k
j
ij
y
1
. Needless to say these estimates can be
obtained by the two local (disjunctive) Kalman
filters, and if necessary, by a smoother. For the sake
of simplicity, let us write the local target estimate
( ) ( )
n
k ik n
y t x
1

after the last observation at

n
t is
processed as
i
x , and in place of
n
t
x , we write the
target state at time
n
t simply as x. Then, in a wide
sense, the track fusion problem is to come up with a
good global estimate x of x from the two local
tracks ( )
n
k k
y
1 1
and ( )
n
k k
y
1 2
. In a narrow sense, the
estimate x is restricted to be a function of only the
two most recent state estimates,
1
x and
2
x ,
calculated only from the two local tracks.
2.2. Bar-Shalom-Campo Fusion Algorithm:
The Bar-Shamon-Campo algorithm as described in
[27] is to calculate the global estimate x from the
two local estimates,
1
x and
2
x , as

2
Simultaneous observation by the two sensors is not
essential to the discussions in this paper but we
assume it for the sake of simplicity.
3
In this paper, positive definiteness is always mean
to be in the strict sense.
2 2 1 1
x W x W x +
(1)
where ( )( )
1
21 12 22 11

+ V V V V V V W
ji jj i
for
{ } 2 , 1 i with i j 3 , with
ii
V being the variance
matrix of the estimate error x x
i
and
ij
V being the
covariance matrix between the estimation errors,
x x
i
and x x
j
.
This estimate can be viewed as a convex
combination of
4
the two estimates,
1
x and
2
x , since
we have I W W +
2 1
. This estimate is a kind of
maximum likelihood estimate in the following
sense. Let ( ) , p be the density function
5
of the joint
distribution of the local estimation errors, x x
1
and
x x
2
, and let ( ) ( ) x x x x p x
2 1
, l be the
likelihood function for the target state x given the
two local estimates,
1
x and
2
x . As shown in [28],
the estimate (1) maximizes this
6
likelihood
function.
2.3. Simple Convex Combination Fusion
Algorithm: It is not clear to the authors to whom
we should attribute this algorithm although it is used
rather commonly probably because of its simplicity
and relatively small amount of necessary data. This
algorithm is also to calculate the global estimate x
by the convex combination, eqn. (1), but with a
simplified weights ( )
1
22 11

+ V V V W
jj i
. In other
words, we obtain this algorithm by ignoring the
covariance matrices,
12
V and
21
V .
2.4 Maximum A Posteriori Probability
Density Estimate: Since our model is linear-
gaussian, the maximum a posteriori probability
density estimate of the target state x given two local

4
The convex combination appears in a pair of
quotation marks since it is not in a usual sense
because of the lack of positivity concept of the
coefficient matrices
i
W .
5
Namely,
( ) { }
2 2 1 1 2 1 2 1
~
,
~
Prob.
~ ~ ~
,
~
x d x x x d x x x d x d x x p .
6
The word likelihood appears within a pair of
quotation marks since the likelihood function here
is not a likelihood function in a usual sense, i.e.,
( ) x x x p
2 1
, .
3
estimates,
1
x and
2
x , is the conditional mean of x
given
1
x and
2
x , which is also the minimal variance
(linear) estimate, and can be written ([31]) as,
( ) ( ) x x L x x L x x + +
2 2 1 1

(2)
where the gain matrix [ ]
1
2 1

z z z x
V V L L L is
calculated from the covariance matrix
z x
V

between
the target state x and the joint local estimates
1
]
1

2
1
def

x
x
z and the variance matrix
1
]
1

22 21
12 11



V V
V V
V
z z
of z . x in (2) is the a priori mean of the target state
x at time
n
t .
2.5 Tracklet Fusion: It might be said that almost
all the algorithms described in a vast volume of the
distributed filtering literature [16] [25] is either
equivalent to or modification of the algorithm of
obtaining the global estimate x from the local
estimates
1
x and
2
x by
x V x V x V x V
1
2
1
22 1
1
11
1

+
(3)
with
1 1
22
1
11
1

+ V V V V . x is the a priori mean
of the state x, and V is its variance matrix of x.
Eqn. (3) can be rewritten as
( ) ( ) x V x V x V x V x V x V
1 1
2
1
22
1
1
1
11
1

+ + ,
and the subtraction x V x V
i ii
1 1

can be viewed as
an operation to extract new information obtained
by sensor i from the mixture with a priori
information. For that reason, this algorithm is called
information differencing in [38]. The so-called
information-matrix form (3) can be rewritten in the
variance-matrix form, which can be viewed as
addition of information in terms of an equivalent
measurement or pseudo measurement to the fusion
agent as described in [20]. The terminology, tracklet,
is attributed to [23] and [24], and is meant to be a
small conditionally independent fragment of a
track.
We should also note that eqn. (3) can be
derived from a general fusion equation expressed as
( ) ( ) ( ) ( ) x p x p x p C x p /
2 1
1
to calculate the
probability density p of the target state x
conditioned by the information from both sensors,
from the two local estimation results represented by
the probability density functions,
1
p and
2
p , and
from the density p of the a priori target state
distribution. As mentioned before, however, the
derivation of this formula (and hence eqn. (3)) is
based on the non-deterministicity assumption.
Nonetheless, usually the non-deterministic dynamics
are used for the time alignment extrapolation. It is
generally understood, therefore, this algorithm works
well only when either the non-deterministicity is
small or the frequency of using the fusion rule (3) is
high enough.
2.6. Comparison of Track Fusion Algorithms:
A very simple example was chosen to compare the
track fusion rules listed above, using a simple two-
dimensional target tracking example. The target
dynamics are defined by
1
]
1

I
I
A
t
0
0
and
1
]
1

qI
B
t
0
where I and 0 are the 2 2 identity and
zero matrices, and and q are two positive
parameters. The initial condition is a zero-mean
gaussian random vector with variance matrix
( ) I I
v
2 2
0
, . Block Diag with
7

2
2
v
q . We
assume a supplementary (or redundant) sensor case
assuming two independent but identical sensors with
[ ] 0 I H
ik
and I R
M ik
2
, and an identical
sampling rate as t t t
k k

+1
with
8
2 n . The
diagonal elements of the variance matrix of the
stationary velocity process are set as
( ) ( )
2 2
2 2 /
M
q
v
. In order to compare the
fusion rules, we vary the process noise intensity q
and the standard deviation
0
of the initial
condition
9
. Fig. 1 shows a result when the process
noise intensity is varied.
In this figure, three fusion rules, (1) convex-
combination, (2) MAP (maximum a priori

7
In this target model, the velocity is modeled by a
stationary stochastic process which is referred to as
Ornstein-Uhrenbeck model [32]. The model is also
called Singer model [33].
8
Namely, each local sensor has only two
measurements.
9
It should be noted that the initial state contributes
to the covariance
12
V between the estimation errors
of the two local estimates as well as the process
noise.
4
probability) fusion, and (3) tracklet fusion, are
compared. Because of the symmetry between the
two sensors, the Bar-Shalom-Campo algorithm is
identical to the simple convex combination
algorithm, and they are simply referred to as the
convex combination algorithm. In this example,
performance measured by the total RMS errors, i.e.,
the square root of the trace of the state estimate error
variance matrix, by any of the fusion rules, does not
deviate much from the best performance, i.e., the
performance achieved by the centralized processing.
For this reason, performance is displayed as the
percent increase in the RMS errors over the minimum
RMS achieved by the centralized tracking.
It is interesting to see that the percentage
increase becomes insignificant when the process
noise is very small as well as when it is very large.
The exception is the performance by the convex
combination algorithm when the process noise is
small. When the process noise is small, the system
becomes almost deterministic, and the system
uncertainty is dominated by the targets initial
condition. The two local processing agents both use
the a priori information. When the local estimates
are fused together, this use of the a priori
information might be double-counted. In the tracklet
algorithm, as well as the MAP algorithm, this double
counting is negated by subtraction, but not by the
convex combination algorithm.
Fig. 2 shows performance of the same three
fusion rules, compared with performance by a
centralized tracking, when the a priori position
variance is varied. The process noise intensity was
set as ( ) 1 / /
3 2
t q
M
. The convex combination
algorithm seems to be consistently worse than other
fusion schemes, which we may attribute to the
double counting of the a priori information. Both
tracklet fusion and MAP fusion schemes deviate from
the optimal performance as the initial state
uncertainty increases.
3. Track-to-Track Association
We will restrict ourselves to the simplest form of
track-to-track association problems, in which two sets
of tracks are to be associated. We are interested in
performance of various track association metrics. To
simplify the problem, we assume that the two sets
have exactly the same number of tracks and that the
true association is always one-to-one. Under a
certain set of assumptions, it can be shown that, given
a two sets of n tracks, the track association problem
can be defined as a problem of choosing a best
hypothesis represented by a permutation a on the set
{ } n ,..., 1 to minimize the association cost,
( )
( )

n
i
i ia
C a J
1
(4)
where each [ ] , 0
ij
C is a given track association
metric.
ij
C results from appropriate
thresholding.
3.1. Problem Statement: We will model targets
as a Poisson-Gaussian finite random set (point
process) in the sense described in [10] [15].
Namely, the number of targets is a Poisson random
variable with a given mean, and given a number of
targets, the target states are independent, and
Fig. 2: Comparison of Fusion Rules with Varying
Initial Uncertainty
( ) /
4
/
2
t
M
Error RMS
Fig. 1: Comparison of Fusion Rules with Varying
Process Noise Intensity
( ) / / Intensity Noise Process
3 2
t q
M

Convex Combination
Tracklet Fusion
MAP Fusion
Convex Combination
Tracklet Fusion
MAP Fusion
Centralized Tracking
(Measurement Fusion)

0
/
M
S.D. Position Priori A
5
identically distributed with a common gaussian
distribution. As a common distribution, we use the
linear-gaussian model described in Section 2.1.
Given N targets, we assume that they are all detected
and correctly correlated together by each local sensor
data processing agent, without any mis-association or
fragmentation, which means that each of the two sets
of local tracks has a one-to-one correspondence to the
set of targets. Using the linear-gaussian model
described in Section 2.1., all the tracks from sensor 1
share the same variance matrix
11
V , and so do those
from sensor 2, the variance matrix
22
V . Thus we can
identify each track i from sensor 1 with the target
state estimate
i
x
1
, and track j from sensor 2 with
j
x
2
. For any pair of track i from sensor 1 and track j
from sensor 2, assuming that they originate from the
same target, the covariance matrix between the two
local target state estimates can be represented by a
single matrix
10
( )
T
V V
21 12
.
3.2. Bar-Shalom Metric: This metric was derived
to be used in a classical chi-square test in [26] and
can be written as
11
( )
2
1
21 12 22 11
2 1

+

V V V V
j i ij
x x C
(5)
3.3. Mahalanobis Metric: This metric is
probably the most frequently used and obtained by
ignoring the covariance ( )
T
V V
21 12
in (5), i.e.,
( )
2
1
22 11
2 1

+

V V
j i ij
x x C
(6)
which can be identified with the square of the
Mahalanobis distance between two gaussian
distributions, ( )
11 1
, V x
i
and ( )
22 2
, V x
j
.
3.4. Chong Metric: This metric can be derived
from a general form of track-to-track association
likelihood ( ) ( ) ( ) ( ) ( ) dx x p x p x p
j i
/
2 1
between

10
By
T
X we mean the transpose of a vector or a
matrix X .
11

A
is a norm on a Euclidean space determined by
a symmetric positive definite matrix A as
Ax x x
T
A
.
track i from sensor 1 having the target state
distribution
i
p
1
described as a density with respect to
a certain measure on a target state space and track
j from sensor 2 with
j
p
2
, where p is the common a
priori target distribution density. In our gaussian
case, we have
2
1
2
1
22
2
2
1
11
1


+
V V
i
V
i ij
x x x x x x C (7)
where x V x V x V x V
1
2
1
22 1
1
11
1

+ with
1 1
22
1
11
1

+ V V V V , and ( ) V x, is the a priori
mean vector and the variance matrix of the target
state x. The subtraction in (7) can be interpreted as
an operation to eliminate the double-counted a priori
information. It can be easily shown that, as V ,
Chong metric (7) converges to Mahalanobis metric
(6).
3.5. Expanded State Metric: This metric can be
obtained by first expanding target state space from
12
d
to
dn
where d is the dimension of the original
target state space, i.e., expanding the target state to be
estimated from ( )
n
t x to ( ) ( )
n
k k
t x
1
, and then by
applying the Chong metric (7) to the expanded target
state estimates. This metric was suggested in [7] but
fully explored in [30]. This expansion of target state
space generally makes calculation directly using (7)
impractical. However, in [30], it was shown that the
metric can be obtained by recursively calculating the
properly defined track likelihood function of each
track to be fused, as well as the fused track, using the
measurements.
3.6. Performance Comparison: Unlike the
performance analysis of Section 2.6., there is no
obvious way
13
of predicting the track association
performance by various association metrics.
Therefore, Monte Carlo analysis was conducted. In
each run, a random set of targets with the average
number 100 of targets was generated according to the
model described in Section 2.6., with the initial
position uncertainty standard deviation being ten

12
( ) reals of set , .
13
In [34], a method for predicting data association
performance is described but it may not be
appropriate when the intensity measure of the random
set (point process) is Gauss-Poisson, rather Uniform-
Poisson.
6
times as big as the measurement error, i.e.,
M
10
0
.
Fig. 3 shows comparison of association
performance by (1) Bar-Shalom metric, (2)
Mahalanobis metric, and (3) expanded state metric,
with various process noise intensity q. For each run,
for each target, it was examined whether the tracks
originating from that target are correctly associated or
not. Then the probability of correct association is
calculated as the number of correctly associated
targets over the total number of targets. Each point in
Fig. 3 was obtained by averaging 300 samples.
Since we have chosen relatively large initial
state variance, we expected the difference between
Chong metric and Mahalanobis metric to be very
small. Consequently, only the simper algorithm, i.e.,
Mahalanobis metric, was evaluated. Because of its
completeness and amount of required computation,
better performance by the expanded state metric is
not surprising at all. However, we should note that
there is no significant difference between Bar-
Shalom metric considering the covariance between
the two tracks and Mahalanobis metric that ignores
such cross-correlation. Lack of difference between
two algorithms may be explained as follows: Bar-
Shalom metric uses the covariance matrix to adjust
the weights in the distance function (5); in a sense,
decreasing with positive correlation and increasing
with negative correlation. However, in our example,
because we assume tracks with uniform quality, those
differences become the same to all the track pairs,
and between two different metrics, the effects appear
only as a scaling difference. This scaling effect
resulted in an extremely small difference caused by
different relative scaling with respect to the fixed
threshold (i.e., the chi-square value of 30).
4. Conclusion
Using very simple linear-Gaussian-Poisson models,
several different track fusion algorithms and track
association metrics were evaluated. Because of
linear-gaussian-ness, track fusion algorithms can be
compared analytically. Track association
performance, however, had to be measured by Monte
Carlo simulations.
Generally computational and
communication requirements were not considered.
Considering these requirements, in practical cases,
those track fusion algorithms requiring evaluation of
cross-correlation between tracks form two sensors
may not be very practical. In many cases, simply
communicating all the measurements in a track with
all the necessary statistics including observable
partials, etc., would be more practical. Under a
reasonable communication restriction, tracklet fusion
would be more practical. Using an appropriate time
intervals between two tracklets from local sites, we
may treat each tracklet as equivalent measurements
from which the expanded track association metric can
be easily calculated.
This approach has been used for a group of
tactical data fusion products called Advanced
Tactical Workstation (ATW [35]) developed by
Advanced C
3
I Systems Unit of Raytheon Systems
Company. As described in [30], in a particular
variant of ATW specially tuned to undersea bearing-
only tracking, the expanded state track association
metric is used very effectively. Recently, it was also
proved, in an ONR-sponsored program, on-board
sensor fusion may be effectively architectured using
this combination of tracklet fusion and expanded
state association metric ([36],[37]).
REFERENCES
[1] Y. Bar-Shalom, and T. E. Fortman, Tracking and
Data Association, Academic Press, Orland, FL 1988.
[2] R. A. Singer, and A.J. Kanyuck, Computer
Control of Multiple Site Correlation, Automatica,
Vol. 7, pp. 455-463, July 1971.
[3] C. Y. Chong, Hierarchical Estimation, Proc.
MIT/ONR Workshop on C
3
, 1979.
[4] C. Y. Chong, and S. Mori, Hierarchical Multi-
Target Tracking and Classification Bayesian
Approach, Proc. 1984 American Control Conf., San
Diego, CA, June 1984.
Fig. 3: Track Association Performance
Expanded State Metric
Mahalanobis Metric
Bar-Shalom Metric
( ) / /
3 2
t q
M
Intensity Noise Process
7
[5] C. Y. Chong, E. Tse, and S. Mori, Distributed
Estimation in Network, Proc. 1983 American
Control Conf., San Francisco, CA, June 1983.
[6] C. Y. Chong, S. Mori, and K. C. Chang,
Information Fusion in Distributed Sensor
Networks, in Proc. 1985 American Control Conf.,
pp. 830 - 835, 1985.
[7] C. Y. Chong, S. Mori, and K. C. Chang,
Distributed Multitarget Multisensor Tracking, in
Multitarget Multisensor Tracking: Advanced
Applications, Y. Bar-Shalom, ed., Chap. 8, Norwood,
MA: Artech House, 1990, pp. 247 - 295.
[8] D. B. Reid, An Algorithm for Tracking Multiple
Targets, IEEE Trans. Autmat. Contr., AC-24, Dec.
1979.
[9] S. Mori, C. Y. Chong, E. Tse, and R. P. Wishner,
Tracking and Classifying Multiple Targets wihout A
Priori Identification, IEEE Trans. Automat. Contr.,
AC-31, May 1998.
[10] S. Mori, Random Sets in Data Fusion Multi-
Object, State-Estimation as a Foundation of Data
Fusion Theory, in Random Sets: Theory and
Applications, ed. By J. Goutsias, R.P.S. Mahler, and
H.T. Nguyen, Springer-Verlag, NY, 1997.
[11] S. Mori, Random Sets in Data Fusion
Problems, Proceedings of 1997 National Symposium
on Sensor and Data Fusion, MIT Lincoln
Laboratory, Lexington, MA, April 1997.
[12] S. Mori, Random Sets in Data Fusion
Problems, Proceedings of 1997 National Symposium
on Sensor and Data Fusion, MIT Lincoln
Laboratory, Lexington, MA, April 1997.
[13] S. Mori, Random Sets in Data Fusion, Proc. of
SPIE Symposium on Data Processing and Tracking
of Small Targets, San Diego, CA, August 1997.
[14] S. Mori, Multi-Target tracking Theory in
Random Set Formalism, Proc. First International
Conference on Multisource-Multisensor Information
Fusion, pp. 116 123, Las Vegas, NV, July 1998.
[15] S. Mori, A Theory of Information Exchanges:
Random-Set Formalism, Proc. 1998 IRIS National
Symposium on Sensor and Data Fusion, Marietta,
GA, April 1998.
[16] J. L. Speyer, Computation and Transmission
Requirements for a Decentralized Linear-Quadratic-
Gaussian Control Problem, IEEE Trans. Automat.
Contr., Vol. AC-24, pp. 266 - 269, April 1979.
[17] A. Willsky, M. Bello, D. Castanon, B. Levy, and
G. Verghese, Combining and Updating of Local
Estimates Along Sets of One-Dimensional Tracks,
IEEE Trans. on Automat. Contr., Vol. AC-27, pp.
799-813, Aug. 1982.
[18] H. R. Hashemipour, S. Roy, and A. J. Laub,
Decentralized Structures for Parallel Kalman
Filtering, IEEE Trans. Automat. Contr., Vol. AC-33,
No. 1, pp. 88 - 93, Jan. 1988.
[19] H. F. Durrant-Whyte, B. S. Y. Rao, and H. Hu,
Toward a Fully Decentralized Architecture for
Multi-Sensor Data Fusion, in Proc. IEEE Int. Conf.
Robotics and Automation, 1990.
[20] B. Belkin, S. L. Anderson, and K. M. Sommar,
The Pseudo-Measurement Approach to Track-To-
Track Data Fusion, in Proc. 1993 Joint Service Data
Fusion Symposium, pp. 519 - 538, 1993.
[21] R. Lobbia and M. Kent, Data Fusion of
Decentralized Tracker Outputs, IEEE Trans. Aero.
Elect. Syst., vol. AES-30, vo. 3, pp. 787-799, July
1994.
[22] O. E. Drummond, Feedback in Track Fusion
without Process Noise, in Proc. SPIE, Signal and
Data Processing of Small Targets, vol. 2561, pp. 369
- 383, 1995.
[23] O. E. Drummond, A Hybrid Sensor Fusion
Algorithm Architecture and Tracklets, in Proc.
SPIE, Signal and Data Processing of Small Targets,
vol. 3163, 1997.
[24] O. E. Drummond, Tracklets and a Hybrid
Fusion with Process Noise, in Proc. SPIE, Signal
and Data Processing of Small Targets, vol. 3163,
1997.
[25] M. D. Miller, O. E. Drummond, and A. J.
Perrella, Tracklets and Covariance Truncation
Options for Theater Missile Tracking, in Proc. 1998
International Conf. on Multisource-Multisensor Data
Fusion (FUSION98), 1998.
[26] Y. Bar-Shalom, On the Track-to-Track
Correlation Problem, IEEE Trans. Automatw.
Contr., AC-25, pp. 802-807, Aug., 1980.
[27] Y. Bar-Shalom, and L. Campo, The Effects of
the Common Process Noise on the Two-Sensor
Fused-Track Covariance, IEEE Trans. on Aero. and
Elect. Syst., Vol. AES-22, No. 6, pp. 803-804, 1986.
[28] K. C. Chang, R. K. Saha, and Y. Bar-Shalom,
On Optimal track-to-Track Fusion, IEEE Trans. on
Aero. and Elect. Syst., Vol. 33, No. 4, pp. 1271-1276,
Oct., 1997.
[29] K. C. Chang, Z. Tian, and R. K. Saha,
Performance Evaluation of Track Fusion with
Information Filter, Proc. First International
Conference on Multisource-Multisensor Information
Fusion, pp. 648 654, Las Vegas, NV, July 1998.
[30] S. Mori, K. A. Demetri, W. H. Barker, and R. N.
Lineback, A Theoretical Foundation of Data Fusion
Generic Track Association Metric -, Technical
Proceedings of Seventh Joint Service Data Fusion
Symposium, pp. 585-594, Laurel, MD, Oct. 1994.
[31] I. B. Rhodes, A Tutorial Introduction to
Estimation and Filtering, IEEE Trans. Automat.
Contr., AC-16, No. 6, pp. 688-706, 1971.
[32] Tiburon Systems, Over-The-Horizon Detection,
Classification and Tracking (OTH/DC&T) System
8
Level Specification Ship Tracking Algorithm, Rev.
C, TIB-01411, June, 1991.
[33] S..S. Blackman, Multiple Target Tracking with
Radar Application, Artech House, Norwood, MA,
1986.
[34] S. Mori, K. C. Chang, and C. Y. Chong,
Performance Analysis of Optimal Data Association
with Applications to Multiple Target Tracking, in
Multitarget-Multisensor Tracking: Applications and
Advances, Vol. II, ed. by Y. Bar-Shalom, Artech
House, 1992.
[35] K. A. Demetri, W. H. Barker, S. Mori, and R. N.
Lineback, Advanced Tactical Workstation,
Proceedings of 1997 National Symposium on Sensor
and Data Fusion, MIT Lincoln Laboratory,
Lexington, MA, April 1997.
[36] W. H. Barker, S. Mori, E. G. Sullinger, and M.
P. Boe, Data Fusion Processing for the Multi-
Spectral Sensor Surveillance System (M4S), Proc.
1998 IRIS National Symposium on Sensor and Data
Fusion, Marietta, GA, April 1998.
[37] W. H. Barker, S. Mori, E. G. Sullinger, and M.
P. Boe, Data Fusion of the Multi-Spectral Sensor
Surveillance System (M4S), Proc. Third NATO/IRIS
Joint Symposium, Quebec City, Quebec, Canada,
Oct., 1998.
[38] J. M. Covino, and B. E. Griffiths, A New
Estimation Architecture for Muti-Sensor Data
Fusion, SPIE Vol. 1478, Sensor Systems for
Guidance and Navigation, pp. 114-125, 1991.