A Track-to-Track Association Algorithm

with Chaotic Neural Network

He Bao-lin, Mao Zheng, Liu Yuan-yuan, Wu Liang

School of Electronic Information and Control Engineering, Beijing University of technology, Beijing, 100124
Phone: 15910925371 Email: [email protected]

offers an inherently parallel computation for solving

Abstract optimization problems. However, the performance of this
approach is limited because the basic Hopfield neural network
A great deal of attentions is currently focused on will always move down in energy space; it will be trapped in
multisensor data fusion. A very important aspect of it is track- the nearest local minima to the initial random state. Some
to-track association and track fusion in distributed technics such as annealing [9], Boltzmann machine [9], have
multisensor-multitarget environments. The approach based on been proposed to solve this problem. But it consume more
Hopfield neural network has been developed. But the time to convergence.
performance of this approach is limited because Hopfield To alleviate the shortcomings of the Hopfield neural
neural network is often trapped in the local minima. This network described above, chaotic dynamics has been
paper try to solve this problem with an approach based on introdued by various researchers and applied to solve
chaotic neural network (CNN). Furthermore, in order to combinatorial optimization problems with encouraging results.
improve the performance of neural network, the association So we try to solve the track-to-track association problem with
statistic between tracks from different sensors is modified. chaotic neural network in this paper. At first, the chaotic
Computer simulation results indicate that this approach is neural network model is present; then, an appropriate energy
more efficient than the algorithm based on continuous function is established for the best association between tracks
Hopfield neural network (CHNN). of different sensors. Consequently, the interconnection
strength between different neurons and input bias of each
Keywords: chaotic neural network (CNN), track-to-track neuron can be obtained. Thus, the track-to-track association is
association, multisensor data fusion. mapped onto the neural network and its output is namely the
answer of track-to-track association. Furthermore, in order to
1. Introduction improve the performance of neural network, the association
statistic is modified.
In recent years, there has been an increasing interest in
At last some computer simulation results are given,
multisensor data fusion for both military and civilian
which indicate that the algorithm of this paper can
applications. Multisensor data fusion combines data from
successfully solve the track-to-track association problem and
multiple sensors system to improve performance and provide
it can be generalized in distributed mutisensor-multitarget
more precise inferences than those to be achieved using a
system.
single sensor system [1]. There are two major types of data
association in mutisensor-multitarget tracking systems [2],
which are measurement-to-track association and track-to- 2. Problem description
track association. In track-to-track association, all the According to the tracking situation, the state space
measured tracks from different sensors are processed in data model of moving targets is given by the following equations:
fusion center to decide whether two or more tracks represent
the same target or not, by some algorithms, such as, statistic X (k + 1) = F (k ) X (k ) + G (k )W (k ) + U (k ) (1)
algorithm [3], Independent sequential algorithm [3] and so on.
However, in the dense target environment or in crossed
Y (k ) = H (k ) X (k ) + V (k ) (2)
tracks situation, some targets can be very close to each other.
The tracks produced by different sensors can confuse track-
where X (k ) is the state vector of the target, Y (k ) is the
to-track association computation algorithms and result in
inaccurate track-to-track association. In this situation, the measurement vector, F ( k ) is the transition matrix, H (k ) is
track-to-track association problem should be considered the measurement matrix, G (k ) is the noise gain matrix,
globally. So a number of other algorithms also have been U (k ) is the forced input, W (k ) is the system noise which is
developed to solve the multisensor-multitarget track-to-track
assumed to be normally distributed with zero mean, V (k ) is
association problem, such as multi-dimension assignment [3],
integer programming [4], Lagrangian relaxation (LR) [3, 5], the measurement noise which is assumed to be normally
and genetic algorithms [3, 6]. The main idea of all these distributed with zero mean, and uncorrelated with W (k ) .
methods is to convert track-to-track association problem into After measurement-to-track association, the state
combinatorial optimization problem, which is difficult to deal estimation of targets is Xˆ i (k | k ) with covariance matrix
s
with.
The approach based on Hopfield neural network [7, 8] Pis ( k ) denotes the estimated state of the is th target detected
has also been developed because neural network technology by s th sensor.
_____________________________
978-1-4244-2732-1/09/$25.00 ©2009 IEEE



Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on January 9, 2010 at 01:39 from IEEE Xplore. Restrictions apply.
 Acording to independent sequential track-to-track In order to solve this 3-D assignment problem, the 3-D
association algorithm [3], the association statistic between Hopfield neural network model has been proposed as an
tracks from different sensors is defined as follwing: inherently parallel method [7]. But it is often trapped in the
nearest local minima to the initial random state. To alleviate
dis−1is (k) = dis−1is (k −1) the shortcoming of the Hopfield neural network, the chaotic
T
(3) neural network model [10,11] and the track-to-track algorithm
+ ⎡⎣ Xˆ is−1 (k) − Xˆ is (k)⎤⎦ Ci−s−11is (k) ⎡⎣ Xˆ is−1 (k) − Xˆ is (k)⎤⎦ based on it is is present in this section.

3.1 Chaotic neural network model

where s = 1, 2, L , M , is = 1, 2, L , N s are the sequence
number of sensors and the sequence number of tracks from The chaotic neural network model we used is to add
sensor s , d i i (0) = 0 and chaotic noise to each neuron of Hopfield neural network and
s −1 s
then gradually decrease it. The network dynamics can be
described as:
C i s - 1 is ( k ) = Pis −1 ( k ) + Pis ( k ) (4)

⎧ N N N

For all the tracks in the common surveillance area, the u

⎪ lmn ( t + 1) = Δ t ( ∑ ∑ ∑ wxyz ,lmn vlmn (t ) + θ lmn )
global statistic can be defined as following: ⎪ x =1 y =1 z =1

⎪ Δt
M ⎨ + (1 − )u lmn (t ) + γη lmn (t ) (9)
τ
bi1i2 L iM ( k ) = ∑d
s=2
is − 1 is (k ) (5) ⎪
⎪ 1
⎪ vlmn (t ) = − ulmn ( t )
A binary event variable is introduced and defined as: ⎩ 1+ e ε

where
⎧1 H0 vlmn (t ) ü output of neuron ( l , m , n ) ,
ρ i i Li ( k ) = ⎨ (6)
1 2 M
⎩0 H1 u lmn (t ) ü internal state of neuron ( l , m , n ) ,
w xyz , lmn üconnection weight between neuron ( x , y , z )
where is = 1, 2, L , N s , s = 1, 2, L , M . H 0 denotes that and neuron ( l , m , n ) ,
i1 , i2 ,L iM th tracks are from the same target, H 1 denotes θ lm n ü input bias of neuron ( l , m , n ) ,
that i1 , i2 , L iM th tracks are from different targets. Then ε ü steepness parameter of the out function, (ε > 0) ,
track-to-track association problem can be represented as γ ü is positive scaling parameter for chaotic noise,
multi-dimension assignment problem. For a data fusion η lmn ( t ) ü the chaotic noise for input to u lmn (t ) .
system of three sensors, the objective of track-to-track
association is met by minimization of the objective function: The chaotic noise can be generated from the logistic map:

N1 N2 N3 η lm n ( t ) = z lm n ( t ) − h (0 < η lm n ( t ) < 1) (10)

∑∑∑ρ
i1 =1 i2 =1 i3 =1
i1i2 i3 ( k ) bi1i2 i3 ( k ) (7)
z lmn ( t + 1) = a ( t ) z lmn ( t )(1 − z lmn ( t ) ) (11)

subject to the constrains:

We let a ( t ) decay exponentially, so that zlmn (t ) is initially
⎧ N1 N2
chaotic and eventually settle to a fixed point z * and
⎪ ∑ ∑ ρ i1i2 i3 ( k ) = 1 for all i3 = 1, 2, L , N 3
⎪ i1 =1 i2 =1 a ( t + 1) = (1 − β ) a ( t ) + β ⋅ a 0 (12)
⎪⎪ N1 N 2
⎨ ∑ ∑ ρ i1i2 i3 ( k ) = 1
(8)
for all i2 = 1, 2, L , N 2
⎪ i1 =1 i3 =1 h = z* = 1 − 1
⎪ N1 N 2 a0
⎪ ∑ ∑ ρ i1i2 i3 ( k ) = 1 for all i1 = 1, 2, L , N 1
⎪⎩ i2 =1 i3 =1 where 1 < a 0 < 2 .9 , 0 < β < 1 .

3.2 Mapping track-totrack association to CNN

3. Track-to-Track association algorithm using chaotic
neural network In mapping track-to-track association to the chaotic
neural network, let the state of vlmn denote an association
states between l th track from sensor 1, m th track from



Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on January 9, 2010 at 01:39 from IEEE Xplore. Restrictions apply.
sensor 2 and n th track from sensor 3 with “1” and “0” Setting scaling parameters A , B , C , D , F properly is
indicating associated and not associated respectively. The very important because we must ensure that the global
energy function of the neural network can be defined as minimum of the energy function correspond to the optimal
solution. Note that some principles for setting scaling
A B parameters [12] can be applied.
E= ∑∑∑
2 l m n
vlmn blmn + ∑ (∑∑ vlmn − 1) 2
2 l m n
4. Simulations
C D
+ ∑ (∑∑ vlmn − 1) 2 + ∑ (∑∑ vlmn − 1) 2 (13)
A number of Monte Carlo studies were carried out to
2 m l n 2 n l m
thoroughly test the modified algorithm in a 2D multisensor
F
+ ∑∑∑ (1 − vlmn )vlmn scenario. All the targets are assumed to be in uniform linear
2 l m n motion in 2D plane (35km × 35 km ), it’s initial position is
assued to be normally distributed, it’s initial velocity and
where blmn is the statistic which is defined by Equ.(5). The initial direction of motion are assumed to be uniformly
first term is the objective function. The second, third and forth distributed within 4 ~ 400 m s and 0 ~ 2π respectively.
terms represent the constraints that each track from one For sensors 1, 2 and 3 the measurement error are assumed to
sensor associated with one and only one track from another be normally distributed with zero mean and standard
sensor. Furthermore, it is enhanced by introducing the fifth deviation σ 1ρ = σ ρ2 = σ ρ3 = 0.15km, σ θ1 = σ θ2 = σ θ3 = 1° . The
term in order to increase the convergence of the neural
coordinates of sensor 1, 2 and 3 are (1 0 km , 6 km ),
network [12]. Derived from Equ.(13), we obtain:
(20 km , 8 km ), (15 km ,10 km ). The state space model of
∂E A moving targets and orther processing is same as reference
= blmn + B ( ∑ ∑ vlmn − 1) [13]. The system noise is assumed to be normally distributed
∂ vlmn 2 m n
(14) with zero mean and the standard deviation
+ C ( ∑ ∑ vlmn − 1) + D ( ∑ ∑ vlmn − 1)
l n l m
Qii = 0.15 × velocity . Kalman filter is used to estamate the
F state vector of the targets. The parameters are set as follows:
+ (1 − 2 vlmn ) A = 30, B = 30, C = 30, D = 30, F = 160 , ε = 0.008 ,
2
γ = 0.1, β = 0.015, a 0 = 1.2, z (0) = 0.1. When k=3,
It is known that the following equation holds [12]: the positions of targets are shown in Fig.1, the evolution of
network’s energy and neurons’ outputs are shown in Fig.2
N N N
∂E and Fig.3 respectively. A number of 100 Monte Carlo runs
∑∑∑w
x =1 y =1 z =1
xyz , lmn vlmn ( t ) + θ lmn = −
∂ vlmn
(15)
were performed in two situations (10 targets and 20 targets)
with this paper’s algorithm and the algorithm based on
Applying Equ.(14), (15) to (9), after the neural network CHNN ( γ = 0 ) respectively. The Proportion of Tracks
converging, the output of the neural network gives the final Correctly Associated of the algorithm based on CHNN ( Pc1 )
solution of track-to-track problem. and this paper’s algorithm ( Pc2 ) is shown in Table 1.
Furthermore, in order to improve the performance of
neural network, the association statistic between tracks from 25
different sensors can be modified as following: From sensor A
From sensor B
From sensor C
⎧d (k ) (if dlm ≤ λ ) 20
Dlm (k ) = ⎨ lm (16)
⎩∞ (otherwise)
15

where λ can be set to be greater than the statistic threshold

y/km

by independent sequential track-to-track association

10
algorithm[3], and we can adjust it according to the actual
situation. Then the global statistic is changed to be:
5
blmn ( k ) = Dlm ( k ) + Dmn ( k ) (17)
k=3
0
By this processing, a lot of global statistics blmn are equal to 5 10 15 20 25 30
infinity, so the active neurons are decreasing sharply in x/km
evolution of the neural network. Thus, the processing above
accelerates the convergence of the neural network. Fig.1. Positions of 20 Targets When k=3.



Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on January 9, 2010 at 01:39 from IEEE Xplore. Restrictions apply.
 7
x 10 network model applied in data fusion system of three sensors
12
is present. Some technics are proposed to improve the
performance of the neural network. It has been shown in
10
Table 1. that the algorithm based on CHN is more efficient
with respect to the algorithm based on continuous Hopfield
8
energy function

neural network. The time complexity of the algorithm is

closer to O ( mN 4 ) . The future work includes studying the
6
relation between parameter setting and scale of neural
4 network, hardware/software co-design.

2 References
[1] J.Llinas, D.Hall, “An introduction to multisensor data fusion”,
0
0 20 40 60 80 Proceedings of the 1998 IEEE International Symposium on
times Circuit and System, vol.6, pp.537-540.
[2] Y.Bar-Shalom, T.Fortman, “Tracking and data association”,
Fig.2. Evolution of the Neural Network’s Energy Academic Press, New York, 1988.
[3] Han Chong-zhao, Zhu Hong-yan, and Duan Zhan-sheng,
1 1 “Multisource information fusion”, Beijing, Tsinghua University
Press, 2006 , pp.330-334.
0.8 0.8 [4] P.R.Williams, “Multiple target estimation using multiple angle-
only sensors”, Hughes Airaraft, Internal Rep., May 1993.
[5] K.R. Pattipati, S. Deb, Y. Bar-Shalom, and R. B. Washburn, “A
neuron output

neuron output

0.6 0.6
new relaxation algorithm and passive sensor data association”,
IEEE Transactions on Automatic Control, vol.37, 1992, pp.
0.4 0.4
198-213.
[6] J.Angus et al., “Genetic algorithms in passive tracking”,
0.2 0.2
Claremont Grad. School, Math Clinic Rep., May 1993.
[7] Tian Bao-guo, He You, and Yang Ri-jie, “Application of
0 0
0 50 100 0 50 100 Artificaial Neural Network in Track Correlation”, Journal of
times times Electronics & Information Technology, Vol.27(2), 2005,
pp.310-313.
Fig.3. Evolutions of the Outputs of Neurons [8] He You, Tian Bao-Guo, “Application of Neural Network in
Track Correlation of Distributed Multisensor Systems”, Acta
Table 1. Proportion of Tracks Correctly Associated Aeronautica et Astronautica Sinica, Vol.25 (3), 2004, pp.300-
Time 10 targets 20 targets 303.
(step) [9] K. A. Smith, “Neural networks for combinatorial optimization:
Pc1 ( 0 0 ) Pc 2 ( 0 0 ) Pc1 ( 0 0 ) Pc 2 ( 0 0 )
A review of more than a decade of research”, INFORMS J.
1 62.1 66.0 56.20 59.70
Comput., vol. 11(1), 1999, pp.15-34.
2 69.0 72.4 61.05 65.10 [10] Terence Kwok, Kate A. Smith, “A Unified Framework for
3 75.2 78.6 67.25 70.05 Chaotic Neural-Network Approaches to Combinatorial
4 79.4 82.3 72.15 75.60 Optimization”, IEEE Trans on neural networks, vol.10.(4),
5 83.2 87.8 77.90 81.25 1999, pp. 978-981.
6 86.7 90.5 80.65 84.95 [11] He Yu-yao, “Chaotic simulated annealing with decaying chaotic
7 88.8 91.8 82.70 87.30 noise”, IEEE Trans. on Neural Networks, vol. 13(6), 2002, pp.
8 89.4 92.4 84.35 88.60 1526-1531.
9 89.5 92.8 85.75 89.05 [12] K.C.Tan, Huajin Tang, and S. S. Ge, “On parameter settings of
10 89.6 93.1 85.60 89.45 Hopfield Networks applied to traveling salesman problems”,
11 89.6 93.0 85.65 89.50 IEEE Transactions on Circuits and Systems-I: Regular Papers,
12 89.7 93.1 85.75 89.45 vol. 52 (5), 2005, pp.999-1002.
[13] He you, Lu Da-jin, Peng Ying-ning, “Two New Track
Correlation Algorithms in a Multisensor Data Fusion System”,
Acta Electronica Sinica
5. Conclusion and future work
In this paper we try to solve track-to-track association
problem with chaotic neural network. The chaotic neural



Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on January 9, 2010 at 01:39 from IEEE Xplore. Restrictions apply.