See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/262270675

IMM-UKF algorithm and IMM-EKF algorithm for

tracking highly maneuverable target: a
comparison

Conference Paper · March 2005

CITATIONS READS

0 72

3 authors, including:

Sebbagh Abdennour Daoud Berkani

Université 8 mai 1945 - Guelma National Polytechnic School of Algiers
7 PUBLICATIONS 9 CITATIONS 100 PUBLICATIONS 248 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Tracking Targets View project

Contourlets transform View project

All content following this page was uploaded by Daoud Berkani on 30 May 2016.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
3URFHHGLQJVRIWKHWK0HGLWHUUDQHDQ&RQIHUHQFHRQ 7
&RQWURO $XWRPDWLRQ-XO\$WKHQV*UHHFH

A Fuzzy IMM-UKF Algorithm for Highly

Maneuvering Multi-Target Visual-based
Tracking
Mohand Saïd DJOUADI Yacine MORSLY Daoud BERKANI
Laboratoire robotique & productique, Laboratoire robotique & productique, Electrical & Computer Engineering,
Ecole militaire polytechnique, Bp :17 Ecole militaire polytechnique, Bp :17 Ecole Nationale Polytechnique, 10, Avenue
Bordj El Bahri 16111 Alger, Algérie. Bordj El Bahri 16111 Alger, Algérie. Hassen Badi, BP.182, 16200 El Harrach
Email: msdjouadi@yahoo.fr Email: ymorsly@yahoo.fr Alger, Algérie.
Email: dberkani@hotmail.com

is about the data association. Effectively, at each sample time,

Abstract— In this paper, we present an efficient filtering
algorithm to perform accurate estimation in jump Markov
nonlinear systems, in case of multi-target tracking. With this
paper, we aim to contribute in solving the problem of model-
based body motion estimation by using data coming from visual
sensors. The Interacting Multiple Model (IMM) algorithm is
specially designed to track accurately targets whose state and/or
measurement (assumed to be linear) models changes during
motion transition. However, when these models are nonlinear,
the IMM algorithm must be modified in order to guarantee an
accurate track. In this paper we propose to avoid the Extended
Kalman filter because of its limitations and substitute it with the
Unscented Kalman filter which seems to be more efficient
especially according to the simulation results obtained with the
nonlinear IMM algorithm (IMM-UKF). To resolve the problem
of data association, we propose to use a Fuzzy approach. The
later is then combined with the IMM-UKF algorithm; the
derived algorithm is called Fuzzy-IMM-UKF.
Index Terms— Estimation, Kalman filtering, Multi-Target
Tracking, Visual servoing, data association.
I. INTRODUCTION
the sensor (camera) present, several measures and
observations, coming from different targets; the problem is
how to affect each measure to the correct target, to resolve this
problem we choose a fuzzy approach. The algorithm derived
from the combination of the later approach and the nonlinear
IMM algorithms is noted Fuzzy-IMM-UKF.
The paper is organized as follows. In section II the
mathematical formulation of 2-D motion is presented. In
section III we briefly present the UKF. We describe in section
IV the nonlinear IMM algorithm UKF based. In section V we
present the Fuzzy-IMM-UKF algorithm. In section VI we
present and discuss the results of simulations. Finally in
section VII we draw the conclusion.
II. MATHEMATICAL FORMULATION OF 2-D MOTION
The mathematical formulation of 2-D motion used is
mainly inspired from Danes, Djouadi, and al in [4]. They
make the hypothesis that the measurements are only the 2-D
Cartesian coordinates of the moving point.
Let s(.) denote the curvilinear abscissa of M over time onto
its trajectory, the origin of curvilinear abscissae is set
arbitrarily. Functions x(.) and y(.), represent the Cartesian

T his paper hope to be a contribution within the field of

visual-based control of robots, especially in visual-based
tracking [3]; tracking maneuvring targets, which may
coordinates of M. The measurement equation may be written
as:
 x (t ) 
themselves be robots, is a complex problem, to ensure a good
track when the target switches abruptly from a motion model

y ( t )
(
 = h s ( t ) , p ( t ) ) (1)
to another is not evident. Because of the complexity and
 
difficulty of the problem, a simple case is considered. The Where p ( .) is a parameter vector function of minimal size.
study is restricted to 2-D motions of a point, whose position is We can see that equation (1) is independent of the type of the
given at sampling instants in terms of its Cartesian motion of M onto its trajectory.
coordinates. This point may be the center of gravity of the The state equation could be written as:
projection of an object into a camera plane, or the result of the
 ( t ) = AX ( t )
X (2)
localisation of a mobile robot moving on a planar ground.
Several of maneuvering targets tracking algorithms are
developed. Among them, the interacting multiple model
 s (t ) 
with X ( t ) =  
(IMM) method based on the optimal Kalman filter, yields  p (t ) 
good performance with efficient computation especially when  
the measurement and state models are linear. However, if the  As 0
latter are nonlinear, the standard Kalman filter should be A equals   , with As the n x n zero matrix with ones
substituted, in our study we choose the recent Unscented 0 0
Kalman Filter (UKF). The other problem treated in this paper, added on its first upper diagonal, and 0 the matrices of
3URFHHGLQJVRIWKHWK0HGLWHUUDQHDQ&RQIHUHQFHRQ 7
&RQWURO $XWRPDWLRQ-XO\$WKHQV*UHHFH

convenient sizes. The continuous time state equation (2) is The output equations are as follows (trajectory parameter are
linear time invariant and independent of M’s trajectory, except considered time-invariant):
on the sizes of s ( .) and p ( .) . Moreover, it may be shown
Straight Line: z (k ) = x(k )  =  d cos α + s (k ) sin α  + v(k ) (3)
∆
that the fundamental matrix F involved its exact discretization  y(k )   d sinα − s(k ) cos α 
at the period T takes the form    
( AT )
i
∆
F = exp ( AT ) = ∑ i =0
n −1
.  s (k ) 
i! Circle:
∆  x ( k )   x0 + R 0 cos R 0  (4)
z ( k ) =  =   + v( k )
The dynamic and measurement noises are supposed to be
 y( k )   y 0 + R 0 sin s( k ) 
stationary, white and Gaussian, non inter-correlated with  R0 
known covariances.
with s(k ) distance covered by the target and v(⋅) zero mean
A. Canonical motion equations Gaussian noise, E {v (k) } = 0, E {v (k) ⋅ v T (k) } = R δ k , k ′
The point M is supposed to move on straight or circular
trajectories at constant or uniformly time-varying speed 2) State Equations
(constant speed or constant acceleration). Those motions
 1 T
belong to the set of the possible behaviours of a non- s(k +1) =  s(k ) + w s (k )
holonomic robot whose wheels are driven at constant Constant velocity  0 1  (5)

velocities or accelerations. p(k + 1) = p(k ) + wp (k )

1) Output equations: One minimal description of a straight
line is defined by the vector p= (α ,d ) shown in figure 1(a),
T
Constant acceleration
which is related to Plucker coordinates. Concerning a circular
trajectory one minimal description is defined by the   T2 
vector p = ( R0, x0 , y0 ) Τ shown in figure 1(b). The origin of  1 T 
  2 
curvilinear abscissa is uniquely defined from those
s( k + 1) =  0 1 T  s ( k ) + w s (k ) (6)
parameterizations.  
 0 0 1 
 
  
p(k + 1) = p(k ) + wp (k )


d s=0
s
α where the random vectors
0 Direction of
(
x(0) = s(0) , p(0)
T T T
) , w(⋅) = (w (⋅) , w (⋅) )
s
T
p
T T
and v(⋅)
M increasing are jointly Gaussian and:
(a) Line E{x(0)} = x̂(0), E{w( k )} = 0, E{v ( k )} = 0,
 x(0) − x̂(0)  x(0) − x̂(0)  T 
s0    
E  w( k )  w( k ′)   = diag (P (0 ), Qδ k ,k ′ , Rδ k ,k ′ )
 v( k )  v ( k ′)  
 
M
 s ( k ) 
and s( k ) =  s( k )  vector of the point M dynamics
y0 + s=0
s(k ) 
0
x0 R0 p(k) trajectory parameters vector

s III. THE UNSCENTED KALMAN FILTER

The basis of the UKF is the unscented transform, where a

Direction of distribution is approximated using a number of vectors which
increasing are passed through the nonlinear function to determine the
(b) Circle probability distribution of the output from the function.
Fig.1. Trajectory Parameterization
3URFHHGLQJVRIWKHWK0HGLWHUUDQHDQ&RQIHUHQFHRQ 7
&RQWURO $XWRPDWLRQ-XO\$WKHQV*UHHFH

A. The Unscented Transform At first the state vector is augmented with the process and
According to Julier and Uhlmann in [5, 6], the unscented noise terms this leads to an na = n + q dimensional vector.
transform is based on the intuition that it is simpler to  x (k ) 
approximate a Gaussian distribution than it is to approximate x a (k ) =  (11)

an arbitrary nonlinear function. A set of vectors, selected to be  w (k )
representative of the probability distribution, are chosen so The process model is now a function of x a ( k ) ,
that their mean and covariance are respectively x and P .
xx
The nonlinear function is applied to each point; the result is a x ( k + 1) = f  x a ( k ) , u ( k ) , k  (12)
set of transformed points with the statistics y and Pyy . The
selection of these vectors is not arbitrarily done but according The unscented transform needs to 2n a +1 vectors which are
to a deterministic algorithm. selected from:
The n-dimensional random variable x with mean and
covariance respectively x and P is approximated by 2n+1
xx  x̂ ( k k ) 
x̂ a ( k k ) =  
weighted points given by:  0 
 q x1 

χ0 = x κ and (13)
W0 =
(n + κ )  Ρ(k k) Ρxw (k k)
Ρa (k / k) =  .
χi = x + ( (n + κ ) Ρ xx ) i Wi =
1 i = 1......n (7) Ρxw (k k) Q(k) 
2(n + κ )
(
χi+n = x − (n + κ )Ρxx ) i Wi + n =
1 The Kalman filter algorithm computes on two principles
phases:
2(n + κ )
- Prediction of the state vector and its covariance matrix
- Estimation of the state vector and its covariance matrix
Where κ ∈R, ( ( n+κ ) Pxx )i is the i th row or column of the by updating the prediction with the current
measurement.
matrix square root of ( n+κ ) Pxx and Wi is the weight which
2n The prediction phase using the unscented transform is as
is associated with the i th point, note that ∑ Wi = 1 .
i =0 follow:
The transformation procedure occurs in three steps:
1. The set of vectors are created by applying equation (7)
1. The transformed set of vectors are: yi = f ( χ i ) (8) to the augmented system given by equation (13)
2n
2. The transformed set of vectors are given by
2. The mean is given by: y = ∑Wi yi (9) χi ( k + 1 k ) = f  χιa ( k k ) , u ( k ) , k  (14)
0
3. The covariance is given by : 3. The predicted mean is calculated by
2 na
x̂ ( k + 1 k ) = ∑ Wi χιa ( k + 1 k )
2n
Ρyy = ∑ Wi {yi − y}{yi − y} Τ (10) (15)
i =0
i =0
4. The predicted covariance is calculated by
According to Julier and Uhlmann [5,6], many interesting 2na
properties could be noticed about this algorithm, for instance it Ρ((k + 1 k ) = ∑Wi {χ i (k + 1 k ) − x̂(k + 1 k )}⋅ (16)
leads to greater accuracy and permit rapid implementation, i =0

thanks to the mean and covariance which are calculated using {χ (k + 1 k) − x̂(k + 1 k)} Τ
i
standard vector and matrix operations. 5. Each prediction vector is instantiate through the
B. The Unscented Kalman Filter measurement model
The unscented Kalman filter is obtained by occurring little
modifications on standard one; this, by restructuring the state zi (k +1k) = h (χi (k +1k),u(k),k) (17)
vector and process and measurement models [5, 6]. 6. The predicted observation is calculated by
Consider the following discrete-time nonlinear dynamical 2na

system: ẑ(k + 1 k ) = ∑Wi z i (k + 1 k ) (18)

i =0
x(k + 1) = f (x (k ), u(k ), k ) + w(k )
z( k ) = h(x(k ), u ( k ), k ) + v( k ) 7. Assuming the fact the measurement noise is additive
and independent, the innovation covariance is
Where x(k+1) ∈ ℜn is the state vector, z(k) ∈ ℜm the Ρ γγ (k + 1k ) = R(k + 1 k ) +
measurement vector, w(k) ∈ ℜq and v(k) ∈ ℜm are Gaussian 2n a
(19)
random vectors with zero mean and covariance matrices Q≥0 ∑W {z (k + 1k ) − ẑ(k + 1 k )}{z (k + 1 k ) − ẑ(k + 1 k )}
Τ
i i i
and R>0, respectively. i =0
3URFHHGLQJVRIWKHWK0HGLWHUUDQHDQ&RQIHUHQFHRQ 7
&RQWURO $XWRPDWLRQ-XO\$WKHQV*UHHFH

8. The cross correlation matrix is calculated by Where:

Ρxz (k +1k) = 1
(20) µ i j ( k − 1 k − 1) = pij µi ( k − 1) i , j = 1,..., r (26)
2na
cj
∑W {χ (k +1k) − x̂(k +1k)}{z (k +1k) − ẑ(k +1k)}
Τ
i i i
i =0 is the probability that model Mi was in effect at k-1 given that
Given these predicted values the state and covariance Mj is in effect at time k conditioned on Zk-1.
estimates are computed according to the equations: And where:

x̂(k +1k +1) = x̂(k +1k) + K(k +1)γ (k +1) (21) - µ i ( k − 1) the probability that the mode Mi is in effect
Τ
Ρ(k +1k +1) = Ρ(k +1k) − K(k +1)Ργγ (k +1k)K (k +1) at time k-1 ,
with γ ( k + 1) the innovation and K(k+1) the weight chosen to r

- c j = ∑ pi , j µ i ( k − 1) the normalizing constants .

minimise the mean squared error of the estimate. i =1

γ (k +1) =z(k +1) −ẑ(k +1k) 2. Model-matched filtering

K(k + 1) = Ρ x z (k + 1 k)Ργγ−1 (k + 1k ) (22)
The above estimate and covariance are used as input to the
filter matched to Mj(k), which uses z(k) to obtain
IV. THE NONLINEAR INTERACTING MULTIPLE MODEL
ALGORITHM x̂
j
(k k ) and P
j
(k k ) .

A. The IMM Algorithm 3. Model probability update

Let a system be described by the equations (notations are
chosen according to [2]) The r model probabilities are updated from the innovation
of the r Kalman filters.
x(k ) = F[ M (k )]x(k − 1) + w [k − 1, M (k )] The likelihood functions corresponding to the r filters are
(23) given by:
z(k ) = H[ M (k )]x(k ) + v [k , M (k )]
Λ j (k ) = Ν {x(k); ẑ j {k / k − 1; x̂ 0j (k − 1 / k − 1)} , (27)
with w (⋅ ) Gaussian process noise and v (⋅) measurement
Gaussian noise.
{
S j k ; Ρ 0j (k − 1 / k − 1) }} j = 1...r
E {w(k) } = 0 , E {w(k) ⋅ w T (k) } = Q δ k , k ′ , 1
µ j ( k ) = Λ j ( k ) cj j = 1, ..., r (28)
{
E {v (k) } = 0, E v (k) ⋅ v (k) = R δ k ,k ′ , T
} c
Where M(k) denotes the model at time k. It’s a finite state r

Markov process tacking values in {M j } , according to a

r where c = ∑ Λ (k )c
j =1
j j is the normalization constant.
j =1

Markov transition probability matrix p assumed to be known,

v and w represent white Gaussian processes and are assumed 4. Estimation and covariance combination.
to be mutually independent.
The output estimates and covariances are computed
A cycle of the IMM algorithm could be summarized in according to the mixture equations
four steps: r
(29)
∑
x̂(k / k ) = x j (k / k )µ (k )
j =1
j

1. The mixed initial condition for filters r

Starting with x̂ ( k − 1 k − 1) , we compute the mixed initial

i Ρ(k / k ) = ∑µ
j =0
j (k )

condition for the filter matched to Mj(k) {Ρ (k / k ) + [x̂ (k / k ) − x̂ (k / k )] [x̂ (k / k ) − x̂ (k / k )] }

j j j Τ

r
x̂0j(k −1/ k −1) = ∑x̂i (k − 1/ k −1)µi j (k − 1/ k −1) (24) The problem here is that this algorithm is designed with the
i =1 assumption that the target motion models are linear, so, for its
i, j =1,...,r optimality, the use of the Kalman filter is recommended.
the covariance corresponding to the above is However, in the case on which the motion models are
nonlinear we propose to use the UKF detailed in section III.
r
The modification to operate on the IMM algorithm is to use as
Ρ 0j ( k − 1 / k − 1) = ∑µ i j ( k − 1 / k − 1) ⋅
state and covariance estimator the UKF.
i =1 (25)
 Thanks to this modification we hope to track accurately targets
i i
[ 0j
 Ρ ( k − 1 / k − 1) + x̂ ( k − 1 / k − 1) − x ( k − 1 / k − 1) ⋅ ] whose models are nonlinear, the resulting algorithm is called

 the Nonlinear Interacting Multiple Model and noted (IMM-
[x̂ i
( k − 1 / k − 1) − x 0j ( k − 1 / k − 1) ] Τ
 i , j = 1 .... r UKF).

3URFHHGLQJVRIWKHWK0HGLWHUUDQHDQ&RQIHUHQFHRQ 7
&RQWURO $XWRPDWLRQ-XO\$WKHQV*UHHFH

V. FUZZY-IMM-UKF ALGORITHM
To resolve the problem of data association, we propose to
use a Fuzzy approach. Fuzzy data association (FDA) is used to
evaluate the association degree between the return and the
actual of interest. Inputs are the measurement innovation.
The association degree is β ijF (k ) = FDA(θ , A, Φ ) ,
where FDA(.) is a fuzzy function based on the set of
membership functions A and the set of propositions Φ [10],
using fuzzy logic methodology, the rules in the FDA rule base
are evaluated. The FDA rules to determine the degree of Association weight
association are provided in table 1, like in [10].
The fuzzy variable, FDA, has 49 If-Then rules; in table 1, the Fig.2. FDA output membership functions
fuzzy term set of input variables are given by
{NL,NM,NS,Z,PS,PM,PL}, where NL, NM NS, Z, PS, PM,
and PL denote Negative Large, Negative Medium, Negative VI. SIMULATIONS AND RESULTS
Small, Zero, Positive Small, Positive Medium, and Positive In this section, we perform some simulations to evaluate
Large, respectively. The fuzzy term set of output variables are our algorithm (Fuzzy-IMM-UKF).
{VL, ML, M, MH, VH}, which denote Very Low, Medium The motion models considered are: - constant velocity on
Low, Medium, Medium High, and Very High, respectively. straight line (M1), -constant acceleration on straight line (M2),
- constant velocity on circle (M3), - constant acceleration on
circle (M4).
X coordinate error
Association degree To explore the capability of our Fuzzy-IMM-UKF
algorithm to track maneuvring targets, various scenarios are
NL NM NS Z PS PM PL considered; among of them we select the typical case of three
NL VL VL ML M ML VL VL highly maneuvring targets with crossing trajectories.
Y
coordinate error NM VL ML ML M ML ML VL
We assume that the target is in a 2-D space and its position is
NS ML ML M MH M ML ML sampled every T=1s. For all cases we assume:
Z M M MH VH MH M M - The measurement noise is zero mean, white, independent
PS ML ML M MH M ML ML of the process noise, and with variance σv2=0.01.
PM VL ML ML M ML ML VL - The process noise is zero mean, white, independent of
PL VL VL ML M ML VL VL the measurement noise, and with variance σw2=0.001.
- The initial condition with the state
For the FDA configurations, the sets of membership X =[s , s , s, α , d , R 0, x0 , y0] is
functions which fuzzify these inputs, defined over appropriate
- The probability transition matrix of four models is
universe of discourse, are shown in Fig.1. The output of the
FDA membership functions, which are used to evaluate the
outputs through the application of fuzzy inference (via max- 0.97 0.01 0.01 0.01
0.01 0.97 0.01 0.01
min rule) and center of gravity defuzzification [11], are shown
p = 
in fig.2. 0.01 0.01 0.97 0.01
 
0.01 0.01 0.01 0.97

- The initial probability of selecting a model is 0.25, that’s

to say, at the start all models have the same chance to be
selected.
- The curvilinear abscissa s (.) remains continuous even if
a trajectory jump occurs.

A. Considered scenario

X and Y coordinate Error We consider that we have to track simultaneously three

maneuvring targets. In order to complicate the scenario, we
Fig.1. FDA input membership functions suppose that the targets follow during there movements,
crossing trajectories.
3URFHHGLQJVRIWKHWK0HGLWHUUDQHDQ&RQIHUHQFHRQ 7
&RQWURO $XWRPDWLRQ-XO\$WKHQV*UHHFH

a) Target 1(black):
The target starts moving according to model M1 until the 1

50th sample when an abrupt trajectory change occur and still

M1
M2
0.9 M3
M4

moving according to this during 50 samples (switching from 0.8

m o de ls pro ba bilities o f ta rge t 3

model M1 to M3). 0.7

0.6

b) Target 2 (blue): 0.5

The target starts moving according to model M1 until the 0.4

th
50 sample when an abrupt trajectory change occur and still 0.3

moving according to this during 50 samples (switching from 0.2

model M1 to M3). 0.1

0
0 20 40 60 80 100 120

c) Target 3 (green): T im e S te p

The target starts moving according to model M1 until the Fig.6. Models Probabilities 3
50th sample when an abrupt acceleration about 0.2 m/s2 occur
and still moving according to this during 50 samples
(switching from model M1 to M2). 25
target 1
40
target 1
target 2 target 2
target 3 target 3
70 35

RM S Po s itio n E rro r in x co o rd ina te

RM S Po s itio n E rro r in y co o rdina te

20

30
60

25
15

50
20
y axis

10
40 15

10
30 5

real t rajec tory 1

5
real t rajec tory 2
real t rajec tory 3
20 es te em ed traj ec tory 1
es te em ed traj ec tory 2
0 0
es te em ed traj ec tory 3 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100

Time Step TimeStep

10
10 15 20 25 30 35 40 45 50

x axis

Fig.3. Real and Esteemed Trajectories Fig.7. RMS x and y position error

4.5 0.05
1
target 1 target 1
M1
target 2 target 2
M2
4 target 3 0.045 target 3
0.9 M3
M4
0.04
0.8 3.5
m o de ls p ro ba bilities o f ta rge t 1

R M S Acc eleratio n Erro r

0.035
0.7
3
R M S S pee d E rro r

0.03
0.6
2.5

0.025
0.5
2
0.02
0.4

1.5
0.015
0.3

1
0.01
0.2

0.5
0.005
0.1

0 0
0 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
0 20 40 60 80 100 120
Time Step
T im e S te p Time Step

Fig.4. Models Probabilities for target 1

Fig.8. RMS Acceleration and Speed Error
1
B. Results interpretation:
M1
M2

Figure 3 shows that the esteemed and the real trajectory for
0.9 M3
M4

the three targets are superposable and almost identical even if

0.8
m o de ls pro ba bilities o f ta rge t 2

an abrupt change occurs on the tracked target dynamic. This

0.7

result is confirmed by the figures (4,5,6,7,8), from this we can

0.6

say that the tracker based IMM-UKF algorithm is a pertinent

0.5

solution to the problem of visual-based tracking highly

0.4

maneuvering targets. In the Other hand figure 1 shows also

0.3

that the data association is correctly done even if the

0.2

trajectories cross each other. This should permit us to say that

0.1

the FDA algorithm computes perfectly and its combination

0
0 20 40 60 80 100 120

T im e S te p
with the IMM-UKF algorithm (Fuzzy-IMM-UKF) would be
Fig.5. Models Probabilities 2
3URFHHGLQJVRIWKHWK0HGLWHUUDQHDQ&RQIHUHQFHRQ 7
&RQWURO $XWRPDWLRQ-XO\$WKHQV*UHHFH

an efficient solution to the problem of highly maneuvering

multi-target visual-based tracking.

VII. CONCLUSION
The model-based body motion estimation by using data
coming from visual sensors still an open problem on which we
try to provide a contribution. In this paper we presented a
nonlinear algorithm which attempts to track efficiently a
highly maneuvering target whose trajectory and/or dynamic
could change abruptly, the algorithm proposed is noted IMM-
UKF. To extend this algorithm to multi-target case, we
combined the later with the FDA algorithm to ensure good
data association.
Simulations show that the Fuzzy-IMM-UKF is a good
investment while we are asked to track a highly maneuvrable
targets whose measurement and/or state models present a
strong nonlinearities, and when there different trajectories
cross each other.

REFERENCES
[1] Y.Bar-Shalom and T.E.Fortman, “Tracking and Data Association,”
Mathematics in Science and Engineering, volume 179, Academic Press,
1988.
[2] Y.Bar-Shalom and X. R. Li, “Estimation and Tracking, Principles,
Techniques and Software,” Artech House, Boston, MA (USA), 1993.
[3] B.Espiau, F.Chaumette, and P.Rives, “ A new approach to visual
servoing in robotics,”IEEE Transactions on Robotics and Automation,
8(3):313-326, June 1992.
[4] P.Danes, M.S.Djouadi, D.Bellot, “A 2-D Point-Wise Motion Estimation
Scheme for Visual-Based Robotic Tasks,” 7th International Symposium
on Intelligent Robotic Systems (SIRS’99), Coimbra (Portugal), 20-23
July 1999, pp.119-128
[5] S.J. Julier, and J.K. Uhlmann, “Unscented filtering and nonlinear
estimation,” Proceedings of IEEE, volume: 92, Issue: 3, March 2004,
pages: 401-422.
[6] S.J. Julier, and J.K. Uhlmann, “A New Extension of the Kalman Filter to
Nonlinear Systems,” In International Symposium on Aerospace/Defense
Sensing, Simulation and Control, Orlando, FL, 1997.

[7] Y.Bar-Shalom and X. R. Li, “Multi-target Multi-sensor Tracking:

Principles and Techniques,” Storrs, CT , YBS Publishing, 1995.

[8] M.S.Djouadi, A.Sebbagh, D.Berkani, ‘‘IMM-UKF and IMM-EKF

Algorithms for Tracking Highly Maneuvring Target,’’ Archive of
Control Sciences, Vol: 1, issue: 1, 2005.
[9] M. Hadzagic, H. Mchalska, A. Jouan, " IMM-JVC and IMM-JPDA for
closely maneuvering targets," IEEE, 2001.
[10] Y.M. Chen, "Anew Data Association Algorithm for Multi-Target
Tracking in a Cluttered Environment," ISIF 2003.
[11] B..P. Singh., W.H. Bailey, " Fuzzy logic applications to multisensor-
multitarget correlation," IEEE Trans. on Aerosp. Electron. Syst., vol. 33,
no. 3, 752-768,1997.

View publication stats