Modeling of Belief Functions For Multi-Target Tracking

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

ICRE’07 university of Bejaia

Modeling of Belief Functions for Multi-Target


Tracking

Abdelmalek MENNAD*, Ahmed DALLIL and Abdelaziz OULDALI


Communications Systems Laboratory
Polytechnic Military School
Algiers, Algeria
*
E-mail: MenadAbdelmalek@yahoo.fr

Abstract—In this paper, we propose a new model reason with uncertainty and suggests a way of
for solving the problem of belief functions combining imperfect data and expressing ignorance.
modelling for data association in multi-target Furthermore, the conflict between the sources to be
tracking (MTT), given a set of environmental combined can be expressed in this theory.
measurements obtained by sensor in a surveillance In this paper we develop a method for MTT with belief
system. The proposed model (PM) exploits belief functions using a model based on a Gaussian function
theory, which is a powerful tool for handling with a special normalisation. The remainder of this
imperfect data. This model (PM) has been tested for paper is organized as follows: section 2 discusses the
a nearly-constant-velocity target and compared to basic notions used in belief theory. In section 3, we
fuzzy set model (FSM) and the Bayesian model formulate the problem discussed in this paper, and give
(BM) in different ambiguous cases and with an overview of the PM for solving the problem of data
different sets of simulated data. The obtained association in MTT, also we present FS and BM
results are very satisfactory and show that this model. The simulation results of comparison between
model provides a useful mechanism for data PM, FS and BM are presented in section 4. Finally in
association. section 6, we conclude this paper.

II. BELIEF THEORY


I. INTRODUCTION
In belief theory, the set of all possible solutions
The belief theory first developed by Shafer [1] (also called hypotheses) of a given problem is called
based on Dempster's previous works [2], is a general frame of discernment, usually denoted by Ω. All the
model of human reasoning under uncertainty. Unlike hypotheses must be exhaustive and mutually exclusive.
Bayesian approach, which is a special case, it doesn't The subsets of Ω, called propositions, can also be
need prior knowledge and it allows for explicit possible solutions to the problem. The set of
representation of ignorance. propositions is the power set of Ω, including Ω itself
In this work we consider the classical problem of and the empty set ∅. Let: Ω = {Hl,H2,…,Hn} and 2Ω
multi-target tracking (MTT); each target is characterized ={∅,H1,H2,H1UH2,…,Ω}. We can define in the power
by its noisy observation. In such case the problem of set 2Ω, a mass function, called also basic belief
observation to track correlation (data association) is the assignment (bba) and denoted by “m”, that associates
one difficulty of MTT, where incoming measurements with every proposition A of 2Ω, degree of belief in
should be correctly assigned to theirs associated track [0,1], such that [6]:
or identified as a new target or false alarm, to obtain ⎧m ( ∅ ) = 0
correct measure for state estimation. The result of data ⎪
m : 2Ω → [ 0.1] ⎨ m ( A) = 1 (1)
⎪ A∑
association is crucial for the overall tracking process.
In the literature, many methods have been proposed to ⎩ ⊆ Ω
deal with the data association problem [3], [4], [5]. In
The bba for a given set A, noted as m(A), represents
recent publications a novel method based on the use of
the belief exactly committed to A and makes no
belief theory was presented [6], [7], [8]. It allows us to
additional claims about any subsets of A.

1
The subset A is a focal element of Ω, if m(A) > 0, As it is used in Transferable Belief Model [9], this
the focal elements and their associated bba define a transformation includes normalization when non-
body of evidence. normalized beliefs are supposed.
From the bba, we can define two other measures of
belief, the belief function and the plausibility function. III. DATA ASSOCIATION BY BELIEF THEORY
The belief function Bel: 2Ω → [0,1] is defined by [7]:
A. Problem formulation
Bel ( A ) = ∑ m ( B) ∀A ⊆ 2Ω (2) In this section, to solve the problem of data
∅≠ B ⊆ A association, we use the method developed in [7],
where the belief theory based data association method
It measures the total belief committed to A. It is equal
is used to solve the problem of vehicle tracking in road
to the sum of all the masses of all subsets B of A. In
situations.
the same way, the plausibility function is defined as:
Let Ωp= {P1, P2, P3,…,PI} and Ωa={A1, A2, A3,…,AJ}
Pl ( A ) = ∑ m ( B) ∀A ⊆ 2Ω (3) be respectively the predicted targets and the validated
( A∩ B )≠∅ measurements at time k.
To study the association relation between these two
The plausibility of A indicates the extent to which sets Ωp and Ωa, a basic idea is to define for each target
one fails to doubt A. It is equal to the sum of all the Pi (i =1.., I) in the set Ωp, J virtual sources of
masses given to sets B whose intersection with information as much as the number of objects in the set
proposition A is different from the empty set. Ωa. Then, each source Sj (j=1…,J) examines if the
Belief theory provides a method for combining the observation Aj is originated from the target Pi or not.
measure of evidence from different sources, using So, we can define J local frames of discernment Ωij
Dempster's rule of combination. This rule combines for each target Pi, related to answering the question: is
two independent bodies of evidence (if we have two the target Pi associated to the observation Aj?
information sources) defined in the same frame of Therefore, the local frame of discernment is composed
discernment Ω, to a new body of evidence. For of two hypotheses: Ωij = {Pi is associated to Aj, Pi is
example, let ml and m2 be two bbas defined in the same not associated to Aj}. Thus, a potential solution for the
frame of discernment Ω. The new body of evidence problem of association between Pi and Aj can be a
noted m12, is given by [6]: proposition from the set:
2Ωij - ∅ ={Pi is associated to Aj, Pi is not associated to
∑ m1 ( B ) m2 ( C ) Aj, I don't know}.
m12 ( A ) = B ∩C = A
(4) As we can see, a potential solution to the problem is
1− ∑ m1 ( B ) m2 ( C ) ignorance, where we cannot decide whether target Pi is
B ∩C =∅ associated to observation Aj or not. The empty set is
The Dempster's rule of combination is commutative not considered since it represents the impossible
and associative, which makes its use very attractive for solution. For simplicity of representation, we noted
the combination of multiple sources since their beliefs respectively the local frame of discernment Ωij and the
can be combined in any order. The basic belief mass power set 2Ωij by:
assigned to the empty set is considered as a measure of
conflict between the combined beliefs [6].
{ }
Ωij = Pi RA j ,Pi RA j and 2
Ωij
{
= ∅ ,Pi RA j ,Pi RA j , Ω a }
Now, having computed the mass, belief and where: R is the abbreviation for association Relation,
plausibility values for all propositions, the decision is { } {
Pi RA j = A j , Pi RA j = A j ," , A j −1 , A j −1 ," , AJ }
made among the different propositions according to
some decision rule. The three most popular decision Since we have J measurements, we obtain J initial
rules are: (i) maximum of plausibility, (ii) maximum beliefs for each target Pi, which are defined over the
of belief, and (iii) maximum of pignistic probability. same frame of discernment. These initial beliefs are
The pignistic probability function, noted BetP, is considered as independent virtual sources of
obtained via a transformation of m, called the information that must be combined to make a decision
pignistic transformation [7]: about the return of the target Pi. This is doing using
Dempster’s rule of combination. This combination
1 m( B) produces a single bba over the power set of this frame
BetP ( A ) = ∑ .
B 1 − m (∅ )
(5) of discernment defined by the set of received
B ⊂ 2Ω ,A⊂ B measurements.
where |B| is the cardinal of B in 2Ω.

2
Table I. Belief matrix ⎧ mij { H1} = m ( i, j )

P1 P2 … PI
⎪⎪ mij { H 2 } = 1 − m ( i, j )
PM → ⎨ (8)
⎪ mij ( ign ) = 0
A1 m1,1 m2,1 … mI,1

A2 m1,2 m2,2 … mI,2 ⎪


⎪⎩ mij ( ∅ ) = 0
…. … … … …
where
AJ m1,J m2,J … mI,J
⎧ H 1 : target i is associated with measurement j
∅ m1,∅ M2,∅ … mI,∅ ⎪
⎪ H 2 : target i isn't associated with measurement j

⎪ign : ignorance of association between i and j
The results of masses combination for all ⎪⎩∅ : the impossible solution
measurements Aj and all targets Pi are summarized and
listed in the table (table I) representing the belief We compare the association performance of PM with
matrix; each element mi;j represent the mass value for FSM and BM.
the proposition: the measurement Aj is originate from
the target Pi. 2) Fuzzy set model “FSM”

Then for decision making, the maximum of belief We choose a fuzzy set model as follows [13]:
is taken as a decision rule i.e. the observation with the ƒ let M max = max ( G( i, j )) and M min = min ( G( i, j ))
j j
highest mass value is chosen as the solution. In the
particular case of the last row: the mass assigned to the M − M min
ƒ let step = max
empty set ∅ due to the fact the virtual sources see 3
(totally or partially) different observations and ƒ let α = Mmin + step and β = Mmin +2 step
representing a conflict between the virtual sources i.e.
ƒ make the following normalisation compared to Mmax:
it reflect partial or total ambiguity in correlation
between the considered target and the measurements. α β G ( i, j )
αnorm = , βnorm = , Gnorm ( i, j ) =
M max M max M max
B. Belief functions modelling ƒ finally the belief function model become:
1) Proposed model “PM”
⎧if Gnorm ( i,j ) ≥ βnorm or Gnorm ( i,j ) ≤ αnorm
The mass function distribution generation is a ⎪
crucial step in establishing the belief model. It is ⎪ ⎧m( H1 ) = Gnorm ( i,j )
⎪ ⎪⎪
closely related to data modelling and the considered ⎪ then ⎨m( H2 ) = 1− Gnorm ( i,j )
application. Numerous algorithms have been proposed ⎪ ⎪
for finding mass function distribution [7], [10], [11], ⎪ ⎪⎩m( ign) = 0
FSM → ⎨ (9)
⎪if αnorm<Gnorm ( i,j ) <βnorm
[12]. In this paper, we have considered the method
suggested in [14]. It takes into consideration data ⎪
inaccuracy and uncertainty. The mass functions are ⎪ ⎧m( H1 ) = 0.25Gnorm ( i,j )
⎪ then ⎪⎪m H = 0.25 1− G
generated as follows:
⎪ ⎨ ( 2) ( norm ( i,j))
ƒ Assuming the Gaussian distribution for the ⎪ ⎪
innovation, the likelihood function Gij is: ⎩ ⎪⎩m( ign) = 0.75
1 1
Gij = exp( − yijT ( Si )−1 yij ) (6) 3) Bayesian model “BM”
2 π Si 2
yij : the innovation vector from observation j to track i, In the Bayesian model, we use the value given in [14]
Si : the innovation covariance matrix for track i. for the development of belief functions model:
⎧ G ( i, j )
ƒ We get the normalisation [15] : ⎪m ( H 1 ) =
⎪ ⎡ ⎤ ⎡ ⎤
m̂ij =
Gij
→ m ( i, j ) = J
ˆ ij
m
(7) ⎪
⎪ ⎢⎣ j
∑ ∑
⎢ G ( i, j ) ⎥ + ⎢ G ( i, j ) ⎥ − G ( i, j )
⎥⎦ ⎢⎣ i ⎥⎦ (10)
max[ Gij ] BM → ⎨
j

j =1
m̂ij ⎪ m ( H 2 ) = 1 − m ( i, j )


ƒ The belief function model is as follows: ⎪ m( ign ) = 0

3
IV. SIMULATION RESULTS To analyse the performances of the PM, we consider
two situations:
The purpose of this section is to demonstrate the
capability of the PM to perform an effective data - Situation 1: three closest parallel targets with initial
association in ambiguous situation, and evaluate its states: (500m,70m/s,500m,80m/s), (550m,70m/s,500m,
performances by comparison with that of FSM and 80m/s) and (600m,70m/s,500m,80m/s). “figure 1-a”.
BM, in term of percent of correct association and - Situation 2: three closest crossing targets at the
precision via RMSE. middle of interval observation; with initial positions
The simulations were done for three kinematics (500m,70m/s,500m,80m/s), (450m,75m/s,550m,75m/s)
targets moving in a two dimensional space with a and (400m,80m/s,598m, 70m/s). “figure 2-a”.
nearly-constant-velocity In this work the extended Kalman filter used to
generate independently for each target state predictions
X (k + 1) = Φ(k ) × X (k ) + Γ(k ) × W (k ) (11)
and estimation is initialized with 70% of the true value.
Where the state vector X(k) at time k consists in When new measurements are available, prediction are
position and velocity components in Cartesian associated with the incoming observations using the
coordinates algorithm explained in the previous section based on
the use of belief theory.
⎡ X 1 (k ) ⎤ ⎡ x(k ) ⎤
⎢ X (k )⎥ ⎢ . ⎥ The simulation results for 100 Monte Carlo run
⎢ 2 ⎥ ⎢ x(k ) ⎥ (12)
X (k ) = = allow us to release the following remarks:
⎢ X 3 (k ) ⎥ ⎢ y (k ) ⎥
⎢ ⎥ ⎢. ⎥ ƒ PM is able to discriminate between targets although
⎣ X 4 ( k ) ⎦ ⎢⎣ y ( k ) ⎥⎦ the two targets are closely spaced and gives good
And performance then FSM and BM, see “figure 1-a,
and figure 2-a”.
⎡1 T 0 0⎤ ƒ From percent of correct association, we show that
⎢ ⎥ the association is better when we use the PM, than
0 1 0 0⎥
Φ (k ) = ⎢⎢ ⎥ (13) if the FSM or BM is used in the two situations
⎢0 0 1 T⎥ (situation 1 and situation 2), see “figure 1-b, and
⎢⎣0 0 0 1 ⎥⎦ figure 2-b”.
ƒ In term of RMSE we can see that PM has the better
And performances then FSM and BM in the two
situations, see “figure 1-c, and figure 2-c”.
⎡T 2 ⎤
⎢ 0 ⎥
⎢ 2 ⎥ V. CONCLUSION
⎢T 0 ⎥
Γ (k ) = ⎢ ⎥ (14) In this paper, a data association method based on
⎢ 0 T2 ⎥ belief theory is introduced, which uses a new model to
⎢ 2 ⎥ determine the mass functions, combines the mass
⎢ ⎥
⎣ 0 T ⎦ functions by using the Dempster rule, finally makes the
decision according to the maximum of belief.
where T = 0.1s denotes the sampling period (the Simulation results have been provided to evaluate the
duration of observation is 200 samples ). W(k) is the association performance of the PM and proved that this
state noise assumed to be Gaussian and centered with model can reliably keep track of the multiple targets
covariance matrix Q = (0.1 m/s2)2 12 without confusion even if they are closely spaced and
Each target is characterized by its noisy position (range even if they are crossing and it shows that the tracking
and bearing) Z(k) given by performances after association by PM are better
compared to those of the FSM and BM. This
⎡ X 2 (k ) + X 2 (k ) ⎤ demonstrates that the PM is simple feasible, efficiency
⎢ 1 3
⎥ and it is well adapted to tracking targets in ambiguous
Z (k ) = ⎢ ⎛ X 3 (k ) ⎞ ⎥ + V (k ) (15)
arctan ⎜ ⎟ association situation.
⎢ ⎜ X (k ) ⎟ ⎥
⎣ ⎝ 1 ⎠ ⎦
Where V(k) is an additive noise, independent from the
noise W(k), with covariance matrix R = diag{(150m)2
(1.5°)2}.

4
(a) (a)
2500 2500

2000 2000

c oordinate (m )
coordinate (m )

1500 1500

track1 track1
1000 track2 1000 track2
track3 track3
observation track1 observation track1
observation track2 observation track2
observation track3 observation track3
500 estimated track1 500 estimated track1
estimated track2 estimated track2
estimated track3 estimated track3

Radar Radar
0 0
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
abscise (m) abscise (m)

(b) (b)
Target1 Target2 Target3 Target1 Target2 Target3
1 0.7 1 1 1 1
PM
PM PM PM
0.95 0.9 FSM 0.9
FSM FSM 0.95 FSM 0.9
0.6 BM
BM BM BM
0.9 0.8 0.8
0.9 0.8
0.85 0.5 0.7 0.7
percent of correct association

percent of correct association

percent of correct association


percent of correct association

percent of correct association

percent of correct association

0.85
0.7
0.8 0.6 0.6
0.4 0.8
0.75 0.6 0.5 0.5

0.3 0.75
0.7 0.4 0.4
0.5
0.7
0.65 0.3 0.3
0.2
0.4
0.65 0.2 0.2
0.6
PM PM
0.1 0.3
0.55 0.6 FSM 0.1 0.1 FSM
BM BM
0.5 0 0.55 0.2 0 0
0 100 200 0 100 200 0 100 200 0 100 200 0 100 200 0 100 200
samples samples samples samples samples samples

(c) (c)
Target1 Target2 Target3 Target1 Target2 Target3
140 160 140 140 140 140
PM PM PM PM PM PM
FSM FSM FSM FSM FSM FSM
BM 140 BM BM BM BM BM
120 120 120 120 120

120
100 100 100 100 100
RMSE Position (m)

RMSE Position (m)

RMSE Position (m)


RMSE Position (m)

RMSE Position (m)

RMSE Position (m)

100
80 80 80 80 80
80

60 60 60 60 60
60

40 40 40 40 40
40

20 20 20 20 20 20
0 100 200 0 100 200 0 100 200 0 100 200 0 100 200 0 100 200
samples samples samples samples samples samples

Figure 1. Tracking of three parallel targets, (a) estimated Figure 2. Tracking of tree crossing targets, (a) estimated
trajectories after data association with PM, (b) percent of trajectories after data association with PM, (b) percent of
correct association when the “PM”, “FSM” and “BM” are correct association when the “PM”, “FSM” and “BM” are
used for data association, (c) RMSE when the “PM”, “FSM” used for data association, (c) RMSE when the “PM”, “FSM”
and “BM” are used for data association. and “BM” are used for data association.

5
REFERENCES [8] E.Blasch, J.Westerkamp, J.Hong, J.Layne, F Garber,
and A. Shaw. Identifying moving HRR signatures with
[1] G. Shafer. Mathematical theory of evidence. Princeton an ATR belief data association filter. In SPIE 2000.
University Press, Princeton, New Jersey, 1976.
[9] Ph. Smets and R. Kennes. The transferable belief model.
[2] A. Dempster. A generalization of bayesian inference. In Intern. J. Intell. Systems, 66(2):191-234, 1994.
Journal of Royal Statistical Society, series B, 30, pages
205-247, 1968.
[10] A. Appriou. Probabilités et incertitude en fusion de
données multisenseurs. revue scientifique et technique
[3] T. E Fortmann, Y Bar-Shalom, and M. Scheffe. de la défense, (11):27-40, 1991.
Multitarget tracking of multiple targets using joint
probabilistic data association. In proceedings of
conference on decision and control, pages 807-812,
[11] T. Denoeux. A k-nearest neighbour classification rule
based on dempster shafer theory. IEEE Transactions
December 1980.
on Systems, Man and Cybernetics, 25(5):804-813,
January 1995.
[4] D. B. Reid. An algorithm for tracking multiple targets.
In proceedings of conference on decision and control,
[12] M. Rombaut and V. Berge Cherfaoui. Decision making
pages 1202-1211, December 1978.
in data fusion using dempster schafer's theory. In 3th
IFAC Symposium on Intelligent Components and
[5] Y. Bar-Shalom and E. Tse. Tracking in a cluttered Instrumentation for Control Applications, Annecy,
environment with probabilistic data association. In France, January 1997.
proceedings of Fourth Symposium on non-linear
estimation theory and its application, pages 13-22, San
Diego, September 1973.
[13] P. Fortemps. Fuzzy sets for modelling and handling
imprecision and flexibility. Polytechnic faculty of
Mons, Belgium, 1996-1997.
[6] N. Megherbi, S. Ambellouis, O. Colot, F. Cabestaing.
Multimodal data association based on the use of belief
functions for multiple target tracking. In 8th
[14] A. Dallil, M. Oussalah and A. Ouldali. Evidential data
association filter, “IPMU 2010”, Dortmund Germany.
International conference on Information Fusion.
Philadelphia 2005.
[15] L. Jang, J. Chao. An information fusion algorithm for
data association in multitarget tracking, IEEE Trans.
[7] A. Dallil, A. Ouldali and M. Oussalah. Data association
Aerospace and Electronic Systems, 1996.
in multi target tracking using belief functions, SETTIT
2009, Hammamet-Tunisia.

You might also like