CODIFICA VIDEO CON MATCHING PURSUIT

1
Pedestrian Tracking Using DCM and Image
Correlation
G.Antonini S.Venegas JP.Thiran and
M.Bierlaire IM2-2004
2
Outline
Introduction ( motivations and objectives ) DCM
for pedestrian dynamic - Pedestrian
behavior modeling (overview) - DCM
specification - DCM calibration - DCM
estimation results Pedestrian detection using
DCM and image correlation Pedestrian tracking
using DCM and image correlation Results
Conclusions and future works
3
Introduction
Motivation new research project conducted with
the aim to integrate state-of-the-art tracking
algorithms with behavioral models for pedestrian
dynamic for video surveillance applications
(IM2.SA).
Objectives our goal is to provide a tool for the
computation of pedestrian trajectories in real,
complex scenarios. These trajectories could then
be used to build statistical density maps and
land-use maps for scene analysis.
4
DCM for pedestrian behavior
5
Pedestrian behavior overview

• Previous approaches are mainly physical-based
  models people as particles (microscopic models)
  subjected to forces people with fluid-like
  properties (macroscopic models, Navier-Stokes or
  Boltzmann-like equations).
• Our approach
• - microscopic model (time-based behavior
  of each pedestrian)
• - walking is a sequence of choices where
  to put the next step? (DCM)
• - dynamical and individual-based spatial
  discretization.

6
Pedestrian behavior overview
time space behavior of individual pedestrians
Two methodologies
Microscopic
Macroscopic
pedestrians with fluid-like properties
Macroscopic
7
The space model
RoI
P
Static pedestrian area
8
Pedestrian behavior our approach
9
The space model
10
10
10
15
Accelerated
20
Constant speed
25
Decelerated
10
Behavioral model
DCM are disaggregate behavioral models designed
to forecast the behavior of individuals in choice
situations
Choice Set
Random variable
Alternatives attributes
dm attributes
11
Behavioral model
Choice Set
Alternatives attributes
dm attributes
Random variable
12
Behavioral model
Choice Set
Alternatives attributes
dm attributes
Random variable
13
Behavioral model
Choice Set
Alternatives attributes
dm attributes
Random variable
14
Behavioral model
Choice Set
Alternatives attributes
dm attributes
Random variable
15
Behavioral model the CNL formulation
Choice Set
Alternatives attributes
dm attributes
Random variable
16
Behavioral model the mixed NL formulation
Choice Set
Alternatives attributes
Socio-economic attributes
Mixed Nested Logit
Random variable
17
Data collection

• Data are collected from a training sequence with
  a frame rate of 10 fps

• We have manually tracked 36 pedestrians using a
  monocular calibrated camera, storing the
  top view positions at each observation.

• We have globally 1675 position observations,
  with a time interval of 3 frames (0.3
  seconds). The observed choice has been measured
  three steps forward in time (0.9 seconds).

• Observations corresponding to static pedestrians
  and to observed choices out of the choice set
  have been removed. Finally, we calibrate the
  model on 1410 observations.

18
Data collection
Trajectory lengths
8 98 pts
(Tanaboriboon,Y.,Hwa,S. and Chor,C.(1986).Pedestri
an characteristic study in Singapore, Journal of
Transportation Engineering 112229-235)
19
looking inside collected data
We manually track 36 pedestrians for a total of
1410 position observations
20
Estimation results CNL
Variable number
Variable name
Coefficients estimates
Asymptotic std error
t-test 0
t-test 1
12
1.3271 e00 1.8426 e-01 7.2024
e00 1.7754 e00
Rho-square - 0.484629
Init log_lik. - 4979.03
Final log_lik. - 2566.05
21
Estimation results mixed NL
Variable number
Variable name
Coefficients estimates
Asymptotic std error
t-test 0
t-test 1
Utility parameters
.. .. ..
Rho-square - 0.313411
Init log_lik. - 4930.08
Final log_lik. - 3384.94
22
Estimation results mixed NL
Variable number
Variable name
Coefficients estimates
Asymptotic std error
t-test 0
t-test 1
Utility parameters
8
1.8573e00 3.8892e-01 4.7757e00
9
- 1.5691e00 5.5359e-01 -
2.8345e00
10
- 1.0134e00 4.8586e-01 -
2.0858e00
11
6.6238e-01 1.8646e-01 3.5523e00
12
5.9938e-01 2.6174e-01 2.2899e00
13
1.0150e00 2.6239e-01 3.8684e00
14
2.6667e00 7.4026e-01 3.6024e00
15
2.5289e00 4.9287e-01 5.1308e00
Model parameters
16
1.4235e00 1.7582e-01 8.0963e00
2.4087e00
Rho-square - 0.313411
Init log_lik. - 4930.08
Final log_lik. - 3384.94
23
Pedestrian simulator
Developed by Mats Weber in the context of the CTI
project SIMBAD
Is initialized with a time-dependent
origin-destination matrix. An itinerary is
associated with each agent.
At each step the utilities and probabilities are
calculated (redhigh utility, bluelow utility)
24
Application of DCM to pedestrian detection and
tracking
25
Pedestrian detection
26
Pedestrian detection
Foreground image
27
Pedestrian detection
28
Pedestrian detection

• Pre-filtering simple thresholding on the
  visual displacements projected on the top-view
  plane. An activation value (starting score) is
  given to each hypthesis. Each bad step consist in
  a penalty.

• Filtering we use the models probabilities to
  give scores to the trajectories over a period of
  T frames

29
Trajectories filtering and detection results
30
Trajectories filtering and detection results
31
Pedestrian tracking
The first approach is to treat tracking as a
sequence of detection cycles deterministic
template matching and behavioral filtering.
32
Pedestrian tracking
We use the model as a prior and the normalized
image correlation as likelihood (at each step the
model is propagated from a MAP estimation on the
previous posterior)
33
Pedestrian tracking
In typical tracking problems, at each frame we
have a model for the targets motion and a
measurement from the image, represented by the
likelihood term.
Normalized correlation between the current
template and the target image

DCM probabilities (at each step the dynamic model
is propagated from a MAP estimation on the
previous posterior)

34
Results
35
Conclusions future works

• DCM are flexible and efficient for pedestrian
  modeling
• The use of behavioral models is usefull both for
  detection and tracking. Can help
• to solve occlusion and illumination condition
  related problems.

36
Future works

• The probabilistic approach to tracking has to be
  improved
• including better representations for the
  posterior distribution
• and the likelihood term (multimodality in the
  correlation distribution).
• We are currently working on a post-clustering
  of trajectories to
• integrate at the end of each detection step.
  Interesting
• preliminary results for pedestrians
  calculation.
• We are working on a scale-adaptive
  head-detector , based on
• statistical properties of edges curvature, to
  use in the
• initialization step and/or for the validation
  of the likelihood
• modes.
• DCM has to be extended to high density
  scenarios with an
• explicit model for fixed and moving obstacles.