Hindawi Publishing Corporation

EURASIP Journal on Advances in Signal Processing

Volume 2010, Article ID 712854, 8 pages
doi:10.1155/2010/712854

Research Article
Vehicle Trajectory Estimation Using Spatio-Temporal MCMC

Yann Goyat,1 Thierry Chateau,2 and Francois Bardet2

1 LCPC, Route de Bouaye, 44341 Bouguenais, France
2 LASMEA, Université Blaise Pascal, 24 Avenue des landais, 63177 Aubière, France

Correspondence should be addressed to Yann Goyat, [email protected]

Received 19 October 2009; Accepted 23 March 2010

Academic Editor: Robert W. Ives

Copyright © 2010 Yann Goyat et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

This paper presents an algorithm for modeling and tracking vehicles in video sequences within one integrated framework. Most
of the solutions are based on sequential methods that make inference according to current information. In contrast, we propose a
deferred logical inference method that makes a decision according to a sequence of observations, thus processing a spatio-temporal
search on the whole trajectory. One of the drawbacks of deferred logical inference methods is that the solution space of hypotheses
grows exponentially related to the depth of observation. Our approach takes into account both the kinematic model of the vehicle
and a driver behavior model in order to reduce the space of the solutions. The resulting proposed state model explains the trajectory
with only 11 parameters. The solution space is then sampled with a Markov Chain Monte Carlo (MCMC) that uses a model-driven
proposal distribution in order to control random walk behavior. We demonstrate our method on real video sequences from which
we have ground truth provided by a RTK GPS (Real-Time Kinematic GPS). Experimental results show that the proposed algorithm
outperforms a sequential inference solution (particle filter).

1. Introduction the applications use sequential methods even though it is not

necessary.
Efficient target tracking is a critical component in many For other situations, deferred tracking is much more
computer vision applications such as visual surveillance appealing, as it is not causal. This allows the optimisa-
or robotics. The object-tracking procedure is intended to tion process to operate over a larger data set (the whole
estimate the state (position, velocity, . . .) of an object at each observation sequence), thus allowing to hope for better
time given an observation sequence. results. Deferred visual multiobject tracking have already
Tracking methods can be divided into two major cat- been successfully experienced on pedestrian tracking in [7]
egories: The first category relates to sequential inference and with a MCMC search in [8].
tracking (also called online or causal tracking), for which the The solution presented in this paper is a spatio-temporal
state of the object at a given time step has been estimated deferred logical inference approach. One of the main
as a function of the record of past and current observations challenges of such methods is that the solution space of
and the record of past states. The second concerns deferred hypotheses grows exponentially related to the (duration)
logical inference (also called offline or noncausal tracking), depth of observation. In the specific case of vehicle tracking,
for which the state estimation at a given point in time uses priors on both driver behavior and the road geometry can be
the entire observation sequence. used. Moreover, the trajectory of the object to be tracked is
Sequential tracking is needed when the tracker’s output driven by a kinematic model. Therefore, we propose an 11-
controls real-time processes, which cannot be delayed (such dimensional reduced state vector of the vehicle trajectory.
as robotic applications). Sequential tracking is also needed Since we use a probabilistic framework, the tracking
when it is not possible to record the observation data, due problem can be seen as the estimation of the distribu-
to its size, or due to regulations. Much work has been done tion of the state vector posterior distribution, given a
on sequential visual tracking (model-based approaches [1– video sequence. We propose a Markov Chain Monte Carlo
3] or learning-based approaches [4–6]). Therefore, most of (MCMC) method to sample the posterior distribution.
2 EURASIP Journal on Advances in Signal Processing

MCMC have been already used in visual tracking. In [9, 10], Priors
a MCMC based particle filter is presented for multiobject
tracking and an extension is proposed to handle a varying X(0) X∗
Proposal
number of objects (Reversible Jump Markov Chain Monte
Video
Carlo, RJMCMC). In [8], the RJMCMC algorithm is used Acceptance sequence
in a deferred logical inference framework to track several rule
vehicles offline from a video sequence.
In MCMC methods, the random walk behavior is driven I0
X(1)
by proposal distributions. We use priors on driver behav- X∗

i=1
p(X | Z) ≈ {X(i) }N
Proposal
ior and road geometry to define efficiently the proposal.

1
k} K
k=
Exploration is achieved with the Metropolis-Hasting rule Acceptance

{I
according to a global likelihood function. rule
We use a likelihood function based on a background
subtraction algorithm. A discrete set of positions of the X(2)
vehicle into the video sequence is generated from the
trajectory state. A generic 3D model of a vehicle is then
projected into each image and then compared to a back-
ground/foreground map of the video sequence. We propose
an efficient implementation of the likelihood function using X(N)
a line integral image to decrease computation time.
Experiments have been done to compare, on real video
sequences, the deferred logical inference approach with a Figure 1: Given a video sequence and an initial state, the method
classic sequential particle filter. samples the posterior distribution of the trajectory using a random
The remaining of the paper is organized as follows. step method (MCMC).
Section 2 presents the probabilistic framework proposed to
solve the tracking problem. Section 3 provides a detailed
description of the vision likelihood function. A set of exper-
a 11-dimensional state space by driver temporal command
imental results along with both qualitative and quantitative
parameters. This method drastically reduces the dimension
analysis is presented in Section 4, before we conclude in
of the state space, thus improving computational efficiency.
Section 5.

2.2. Driver Command and Vehicle Priors. The driver com-

2. Proposed Method mands are the steering wheel angle, and the vehicle longitu-
dinal acceleration, from which we deduce the vehicle speed
This section describes the core of the method, based on a
through integration. The experiments presented below have
probabilistic framework. Figure 1 represents an illustration
been conducted on a mid-velocity curve. While traveling
of the algorithm. Given a video sequence and an initial
such a curve, a light vehicle driver’s command law is
state, the method samples the posterior distribution of the
commonly modelled by a trapezoid, with steering wheel
trajectory using a random walk method (MCMC). In the
angle velocities lying between 1.5 and 4 degrees per second,
following, we begin by presenting the state vector associated
and with absolute longitudinal accelerations lying between
to the trajectory model. Then, we give an overview of the
1 m·s−2 and 3 m·s−2 . In order to take into account a
Monte Carlo Markov Chain algorithm used to sample the
wider range of driver commands, and the steering system
posterior distribution. Finally, we show how to generate
nonlinearities due to frictions and mechanical compliances,
new proposals by sampling from an object-specific proposal
we use a more compliant model: we model the steering
distribution.
command with a double sigmoid (one for entering the curve,
and one for releasing from it).
2.1. State Vector Reduction. In a spatio-temporal deferred As the experiments presented below have been conducted
logical inference approach, the solution space of hypotheses on the second half of the curve, a single sigmoid is used to
grows exponentially related to the number of observation define the steering angle generator, from parameters defined
frames. Estimating a single vehicle planar trajectory along in Section 2.2 (cf. Figure 2)
a 100 frame sequence, involves a state space of dimension
300 (planar position and orientation are estimated for each . θ
fδ (θ δ , k) =   δ,2   + θδ,1 ,
frame). Conducting a Monte-Carlo search in such a space 1 + exp θδ,3 θδ,4 − k / θδ,2  (1)
is computationally intractable! To avoid this problem, we
do not consider the vehicle position sequence as the state k = time.
vector, but we implement a trajectory generator, lying on
driver behavior priors, road geometry priors, and vehicle The same reasoning applies to the vehicle speed generator
kinematic priors. The vehicle trajectory generator detailed fv (θ v , k), calculated as fδ (θ δ , k), swapping index δ into index
in the following, generates trajectory samples, defined in v.
EURASIP Journal on Advances in Signal Processing 3

Input: The first element of the chain X0 and its weight

proportional to its likelihood: π(X0 ) ∝ P(Z | X0 )
for n = 1 to N + NB do
θδ,1 + θδ,2

θδ,3 - Choose a move m ∈ {1, . . . , M } among all the parameter

of the state X according to prior q (m).
- Draw a proposal X∗ from the distribution q(X∗ | X)
with X = Xn−1
- Evaluate its joint likelihood: p(Z | X∗ )
θδ,4 k - Compute the acceptance ratio using Metropolis-Hasting
θδ,1
rule:
Figure 2: Graphic representation of the sigmoid parameters. p(Z | X∗ ) q(X | X∗ )
α = min 1, ×
p(Z | X) q(X∗ | X)
- Add a nth element to the chain Xn = X∗ with
probability α, (otherwise Xn = Xn−1 ).
To model the vehicle, we use a plain kinematic model, as
end for
described in Section 3.1. This model allows us to iteratively Burn-in: delete the NB first elements of the chain.
generate xk , yk , and αk , for every time step k. Output: N-element Markov Chain of state hypothesis:
The vehicle trajectory generator is represented by a {Xn }n=NB +1,...,NB +N
.
random state vector X = (l, θ δ , θ v )T with
.
(i) l = (x0 , y0 , α0 ) represents the initial position and Algorithm 1: MCMC algorithm.
orientation of the vehicle (into a world reference
frame),
.
(ii) θ δ = (θδ,1 , . . . , θδ,4 ) are the parameters of a sigmoid
function δk = fδ (θ δ , k) representing the discrete 2.4. Proposals. At iteration n, the MCMC generates a new
temporal evolution of the steering angle, proposal by sampling from a proposal distribution q(X∗ |
.
(iii) θ v = (θv,1 , . . . , θv,4 ) are the parameters of a sigmoid X(n−1) ) defined by
function vk = fv (θ v , k) representing the discrete
 
temporal evolution of the vehicle velocity. q X∗ | X(n−1) = q (m)q X∗ | X(n−1) , m , (5)
m∈{1;...;M }
2.3. MCMC. We want to estimate p(X | Z), the posterior
probability density for a model’s parameters X, given some
observed data Z. Monte-Carlo methods assume that the where q (m) is a prior distribution used to select the
posterior distribution can be approximated by a set of N parameter index of X to be modified (M denotes the size
samples: of X). A parameter-specific proposal distribution is then
 N
defined by
p(X | Z) ≈ X(n) . (2)
n=1
.  
−1)
Sampling from p(X | Z) is a hard problem and many q(X∗ | X, m) = p Xm∗ | Xm(n−1) δ X ∗j − X (n
j . (6)
methods have been proposed. Metropolis-Hasting is a ran- j=
/m
dom walk algorithm designed to approximate a stationary
distribution. At each step, a state X∗ is proposed according Here, only the mth component (m is selecting with the prior
to a proposal density q(X∗ | X). The proposal state is then distribution q (m)) of the state vector is moved at iteration
accepted or rejected according to an acceptance ratio defined n; the other parameters remain unchanged. The MCMC is
by the Metropolis-Hasting rule summarized in Algorithm 1.
p(Z | X∗ ) q(X | X∗ )
α = min 1, × . (3)
p(Z | X) q(X∗ | X) 3. Observation
Metropolis-Hasting rule can be used to build a Markov
This section presents the observation function defined to
Chain which approximates the posterior distribution p(X |
compute the likelihood p(Z | X = X(n) ) probability to
Z). The resulting method is called Markov Chain Monte
observe the video sequence, given a sample X(n) . Figure 3
Carlo. Moreover, the Nb first elements of the chain are
illustrates the observation process. A discrete set of positions
removed (burn-in) in the final sampling set. An estimate of
of the vehicle into the video sequence is generated from
the state is given by a maximum likelihood rule applied to
the trajectory sample X. A generic 3D model of a vehicle
the particle set
is then projected into each image and compared to a
.
N  background/foreground map of the video sequence.
X = arg max δ X − X(n) , (4) We propose an efficient implementation of the likelihood
X
n=1
function using a line integral image to decrease computation
where δ is the Dirac function. time.
4 EURASIP Journal on Advances in Signal Processing

Kinematic L L
modelδ State Vedio
vector
Z
X
Car
Bicycle δ
kinematic model
Camera
parameters
Figure 4: The bicycle model synthesizes the displacement of a
four-wheel vehicle, through the displacement of two wheels whose
centers are connected by a rigid axis of length L. Ackerman’s theory
serves to estimate the steering angle of the front axis of a vehicle
traveling at low speed.

k
k z0
{xk }K
k=1 {Ik }K
k=1

(Rw ) T n θnt
(R0 ) (xt ) R0 y0

K  zw ynt x0
{ p(zk | xk )}K
k=1 p(Zk | Xk ) = k=1 p(zk | xk )
yw
Rw
xnt
xw
Figure 3: Illustration of the likelihood function. A discrete set of
positions of the vehicle into the video sequence is generated from Figure 5: Example of a simple three-dimensional geometric model
the trajectory sample. A generic 3D model of a vehicle is then pro- used for a vehicle. It is composed of two cubes. The coordinate
jected into each image and compared to a background/foreground system associated with the cube and the other system associated
map of the video sequence. with the scene are related according to pure translation. The plane
(Oxy) of the world coordinate system and component axes are
merged with the GPS coordinate system.
3.1. Building a Discrete Set of Vehicle Positions. Let X define
a discrete set of temporal positions and orientations of the
vehicle, associated to a sample X of the posterior distribution 3.2. Computing p(zk | xk ). Since the video sequence comes
from a static camera, vehicle extraction is achieved using
.
X = {xk }Kk=1 , (7) a background/foreground extraction approach. We use a
nonparametric method [11], based on discrete modelization
. of the background probability density of the pixel color
with xk = (xk , yk , αk )T is a vector which gives the position .
and orientation of the vehicle at time k into a world reference (RGB). The algorithm provides a set of binary images I =
. T
frame Rw associated to a planar ground. xk can be computed {Ik }K
k=1 , where Ik (u) = 1 if the pixel u = (ux , u y ) is
in a recursive way using a simple kinematic model of the associated to foreground and Ik (u) = −1 if the pixel is
vehicle. Here, we used a bicycle model (cf. Figure 4) associated to background.
A simplified three-dimensional geometric model of
xk = xk−1 + T · vk−1 · cos(αk−1 ), the vehicle is used, as depicted in Figure 5. This model
is composed of two nested parallelepipeds. In a general
yk = yk−1 + T · vk−1 · sin(αk−1 ), (8) case, the model may be more complex and contain PM
v
αk = αk−1 + T · k−1 · tan(δk−1 ), parallelepipeds. Let M(R0 ) = {Mi(R0 ) }i=1,...,NM represent the
L model’s set of cube vertices (NM = 8 × PM ), expressed
where T is the sample time used for the video acquisition within a coordinate system associated with model R0 . This
and L denotes the wheelbase (distance between front and rear coordinate system is selected such that the 3 axes all lie in the
wheels). δk and vk are given by the steering angle and velocity same direction as that of the world coordinate system Rw .
parametric functions presented into Section 2.2. Each point of the vehicle model is projected onto the
The likelihood function p(Z | X) can be written by image via the following equation

p(Z | X) = p(z1 ; z2 ; . . . ; zK | x1 ; x2 ; . . . ; xK ) m (R

 i ∝ Cc ·(Rw ) T(R0 ) (xk ) · M 0)
, (11)
(9) i

and assuming independence of random variables with M homogeneous coordinates associated with point
M; Cc is the camera projection matrix, and (Rw ) T(R0 ) (xk )

K the homogeneous transformation matrix between the world
p(Z | X) = p(zk | xk ). (10) coordinate system and the system associated with the 3D
k=1 model (cf. Figure 5).
EURASIP Journal on Advances in Signal Processing 5

The set M(Ri ) = {mi }i=1,...,NM is thus built based on the x


projection of 3D model points within the image. Ik
Ik Convex hull
For a given position xk , the likelihood is linked to
the difference between the number of foreground and y
background pixels inside the vehicle model projection in the
image. This computation performed for each particle spends
SIGNIFICANT processing time, and we are proposing
 
herein a fast likelihood calculation method based on an
Ik ((x1 , y1 )T ) Ik ((x2 , y1 )T )
approximation of the 3D model projection in the image
through its convex hull. Foreground binary map Line-integral image
.
Let E (M(R0 ) ; xk ) = {ei }i=1,...,Ne (ei = (xie , yie ) as Figure 6: Illustration of the vision likelihood computation. The
coordinates of ei in the image plane) be the list of convex 3D model of the vehicle (shown in green/clear) is reprojected onto
hull points. (Calculation of the convex hull is not developed the image generated from the background-shape extraction. This
in this article; the calculation procedure is conducted using projection is approximated by its convex hull (shown in red/dark
a classical algorithm with a complexity expressed in O(N · on the right image). The likelihood calculation proceeds in a line-
.
log N).) We will now define Ek = E (M(R0 ) ; xk ) in order by-line integral image of the log-likelihood ratio.
to streamline notations. The likelihood calculation may be
performed efficiently by use of a line-by-line integral image
defined by 4.1. Experimental Details
 T  x
 T  4.1.1. Initialisation. The first sample of the MCMC must
IΣk x, y = Ik i, y . (12)
i=1
be initialized using priors. We use a data driven method
to compute the initial position of the vehicle on the road
Points ei are categorized by pairs featuring the same y- (x0 , y0 ). A nonparametric blob detector [11] is applied to
coordinate values, such that the background/foreground image I0 . The initial velocity is
 provided by a specific sensor. Other parameters are initialized
       
Ek = x1e , y e , x2e , y e , x3e , y e + 1 , x4e , y e + 1 , . . . using priors given by vehicle or driver behavior. The dimen-
(13) sions of the geometric model are defined for each vehicle
   
N N with a stochastic process on width and length parameters
xNe −1 , y + e , xNe , y e + e
e
. and using the likelihood computation (cf. Section 3.2). If
2 2
dimensions seem to be incoherent, a standard vehicle is
Convex hull coding within the set Ek necessitates a chosen.
shape discretization along the image lines. Moreover, special
attention needs to be paid to coding the upper and lower
4.1.2. Proposals. A key point of the method concerns the
extremities. On the other hand, it is not at all necessary
control of the random walk behavior using proposal distri-
to sort points positioned on the same line. A compliance
butions. Parameter-specific proposals are defined. Since both
measurement relative to a convex hull is computing from the
lower and upper bounds can be defined for all parameters, we
integral image by application of the following relation:
choose proposals according to a Beta distribution
Ne /2      ξ1 −1  
∗ ξ2 −1
a(Ek ) = 2 · IΣk e2 j − IΣk e2 j −1 − x2e j − x2e j −1 . P Xm∗ | Xm(n) ∝ Xm∗ · 1 − Xm (16)
j =1
(14) parameters ξ1 and ξ2 are computed such as the maximum of
the distribution is obtained for Xm∗ = Xm(n) .
Figure 6 describes the principle behind the likelihood
calculation method using the integral image. A line-by-line
scanning is performed as part of this method. 4.1.3. Details about the Sequential Method. Behavior of the
Finally, the likelihood expression is written by proposed method is compared with a sequential particle
.
filter. The state vector is defined as Xk = (x y , yx , αx , vx , δx )T .
p(zk | xk ) ∝ CE−k1 max(0, a(Ek )) (15) Dynamics are controlled by the kinematic bicycle model with
a zero centered normal law applied to both the steering angle
. 
with the normalization constant CEk = Nj =e /2 e e
1 |x2 j − x2 j −1 | and velocity variation. Moreover, the likelihood function
defining the surface of the convex hull. is slightly modified by removing the normalizing constant.
The particle set is resampled at each iteration using a SIR
algorithm.
4. Experimental Validation
In this section, experimental results are presented to high- 4.2. Results. In order to compare the two methods, a vehicle,
light the relevance of our tracker. We compare the Offline equipped with a RTK GPS accurate to within one centimeter,
proposed approach to a sequential stochastic filter (particle has been used. (A calibration between the GPS reference
filter). frame and the camera has been achieved but details are
6 EURASIP Journal on Advances in Signal Processing

Table 1: Position error (the true position is given by a RTK GPS) for the proposed deferred logical inference method and a sequential particle
filter.
Method Position error (m) Position std. (m) Orientation error (degrees) Orientation std. (degrees)
Sequential filter 0.27 0.26 3.67 3.36
Deferred logical inference 0.20 0.22 1.12 0.97

100 100

80 80
Accuracy (%)

Accuracy (%)
60 60

40 40

20 20

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Position tolerance (cm) Orientation (yaw) tolerance (degrees/10)

Deferred Deferred
Sequential Sequential
(a) (b)

Figure 7: Percentage of correct position/orientation related to the tolerance. (a) position absolute tolerance. (b) orientation (yaw angle)
absolute tolerance.

not presented in this paper.) This vehicle traveled through trajectory, thus bringing more time consistency than the
the test section 20 times along various trajectories at speeds sequential method.
ranging from 40 to 80 km/hr. The error was quantified as the Figure 8 illustrates the two methods on a real sequence.
average distance between each estimated vehicle position and Curves on the right column show zooms on local trajectories.
the straight line passing through the two closest GPS points. The middle column illustrates the image projection of the
For each test, at least five vehicle runs were carried out, vehicle position for the sequential method. The right column
which enabled deriving a very rough statistic on the recorded illustrates the image projection of the vehicle position for
measurements. For the tests actually conducted, the vehicle the deferred method. It is of high interest to notice the
has been tracked in a curve over a distance of approximately noisy estimation provided by the sequential method, where
100 m (minimum radius = 130 m). the estimated trajectory does not seem to match the vehicle
All the experiments presented here have been done using kinematic model. The reason for this weak consistency is
200 particles for the two methods. that the maximum a posteriori estimate may be found
Table 1 presents the average error and related standard on different particles at every time step. In contrast, the
deviations for the two tested methods. The proposed spatio-temporal deferred approach ensures faithfulness to
deferred logical inference provides a lower global error than the model, thus explaining the observed improvement.
the sequential particle filter.
Figure 7 plots the estimation accuracy as a percentage 5. Conclusion
of correct positions (vertical axis) versus an error tolerance
(horizontal axis) for both methods. On the left graph the We have presented a solution for estimating vehicle tra-
error tolerance is the position absolute error, ranging from jectories using a single static color camera. A spatio-
0 to 50 cm, while on the right graph the error tolerance is temporal deferred logical inference solution which takes
the vehicle orientation (yaw angle) absolute error, ranging into account both vehicle kinematics and driver behavior
from 0 to 5 degrees. The curve associated to the proposed has been proposed, using a stochastic approach to estimate
method outperforms the sequential particle filter both for the the posterior distribution of the trajectory. By choosing a
position and the orientation estimation. Moreover, the right MCMC, the random walk evolution is controlled by injecting
graph emphasizes the benefit of the deferred method, which priors on both driver and vehicle behavior and on geometric
integrates the vehicle and driver priors in every generated knowledge about the road. Moreover, a global likelihood
EURASIP Journal on Advances in Signal Processing 7

177

176

175

174

115.5 116 116.5 117

198
197
196
195
194
193
192
191
190
189
103 104 105 106 107 108 109 110 111

206.8
206.6
206.4
206.2
206
205.8
205.6
205.4
205.2
205
204.8
99.2 99.6 100 100.4 100.8

Real trajectory
Sequential method
Deferred method

Figure 8: Snapshots illustrating the two methods. Left column: zoom on local trajectories. Middle column: the bounding box illustrates
the position of the vehicle estimated with the sequential method. Right column: the bounding box illustrates the position of the vehicle
estimated with the sequential method.

function using background/foreground binary extraction the entire curve, the system is composed of three color
has been proposed, with an efficient implementation. cameras with very little overlap. The system has successfully
Experiments have been achieved to demonstrate that the analyzed observations recorded under actual traffic condi-
proposed method outperforms a classic sequential particle tions over several-day periods.
filter solution using statistics performed on real video
sequences. Two points explain this improvement. First, the
spatio-temporal deferred approach processes over the whole
data set, thus ensuring time consistency. Second, the spatio- References
temporal deferred approach, unlike the sequential approach, [1] D. Comaniciu, V. Ramesh, and P. Meer, “Real-time tracking
ensures total faithfulness to the model at any time step, of non-rigid objects using mean shift,” in Proceedings of IEEE
because the maximum a posteriori estimate may be found Computer Society Conference on Computer Vision and Pattern
on different particles at every time step. Recognition, vol. 2, pp. 142–149, 2000.
The method discussed in this paper is currently operating [2] M. Isard and A. Blake, “Condensation—conditional density
24 hours a day with various weather conditions to provide propagation for visual tracking,” International Journal of
statistics on curve trajectories. For the purpose of covering Computer Vision, vol. 29, no. 1, pp. 5–28, 1998.
8 EURASIP Journal on Advances in Signal Processing

[3] M. Isard and J. MacCormick, “BraMBLe: a Bayesian multiple-

blob tracker,” in Proceedings of the IEEE International Confer-
ence on Computer Vision, vol. 2, pp. 34–41, 2001.
[4] O. Williams, A. Blake, and R. Cipolla, “A sparse probabilistic
learning algorithm for real-time tracking,” in Proceedings of the
IEEE International Conference on Computer Vision, vol. 1, pp.
353–360, Nice, France, 2003.
[5] S. Avidan, “Support vector tracking,” in Proceedings of the IEEE
Computer Society Conference on Computer Vision and Pattern
Recognition (CVPR ’01), vol. 1, pp. 184–191, Kauai, Hawaii,
USA, 2001.
[6] S. Avidan, “Ensemble tracking,” in Proceedings of the IEEE
Computer Society Conference on Computer Vision and Pattern
Recognition (CVPR ’05), vol. 2, pp. 494–501, IEEE Computer
Society, Washington, DC, USA, 2005.
[7] F. Fleuret, J. Berclaz, R. Lengagne, and P. Fua, “Multicamera
people tracking with a probabilistic occupancy map,” IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol.
30, no. 2, pp. 267–282, 2008.
[8] Q. Yu, G. Medioni, and I. Cohen, “Multiple target tracking
using spatio-temporal Markov chain Monte Carlo data associ-
ation,” in Proceedings of the IEEE Computer Society Conference
on Computer Vision and Pattern Recognition (CVPR ’07), 2007.
[9] Z. Khan, T. Balch, and F. Dellaert, “MCMC-based particle
filtering for tracking a variable number of interacting targets,”
IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 27, no. 11, pp. 1805–1819, 2005.
[10] K. Smith, D. Gatica-Perez, and J.-M. Odobez, “Using particles
to track varying numbers of interacting people,” in Proceedings
of the IEEE Computer Society Conference on Computer Vision
and Pattern Recognition (CVPR ’05), vol. 1, pp. 962–969, 2005.
[11] Y. Goyat, T. Chateau, L. Malaterre, and L. Trassoudaine,
“Vehicle trajectories evaluation by static video sensors,” in Pro-
ceedings of the 9th IEEE International Conference on Intelligent
Transportation Systems (ITSC ’06), pp. 864–869, 2006.