ModelBased Object Tracking in Road Traffic Scenes

1
Model-Based Object Tracking in Road Traffic
Scenes

• Presented by
• Feng Tang
• May, 11, 2005

2
Goal

• Robustly track vehicles under arbitrary viewing
  angle changes
• Make use of priori knowledge for tracking

3
Major novelties

• 3D vehicle model
• structure and motion
• Illumination model
• verify the hypotheses for object candidates

4
The model priors

• Structure prior
• Motion prior

5
The 3D vehicle model

• 12 parameters

6
Generative model
7
The motion model
Translational velocity

• Constant velocity

Angular velocity
Object center
Orientation angle
8
Primitive features
Notations Line segment attribute

• Edge segments
• Model edges
• Data edges

9
Matching primitives

• Distance measure Mahalanobis distance

10
The matching
Model segment set
Data segment set
Model interpretation
11
Model Initialization

• Goal Compute the
• Steps
• Motion segmentation
• Motion vector clustering
• Back-projection ( need calibration )

12
3D-2D Alignment

• Objective function
• The weighted sum edge segment matching residuals
• Levenberg-Marquardt minimization

13
Another prior illumination model

• Why?
• Shadow distraction
• How?
• Inclusion of the shadow edges into the modeling
• Parallel light

14
The scene structure
15
Recursive motion estimation

• State vector

• State transition

16
The measurements

• Stack vector from matched segments

17
The MAP formulation
Which is equivalent to
18
The solution

• Levenberg-Marquardt method
• Equivalent to IEKF( iterated extended Kalman
  filter)

19
Overview
20
Experiments Leaving car
21
(No Transcript)
22
Experiments parking
23
Experiments incoming
24
(No Transcript)
25
Experiments mulilane intersection
26
(No Transcript)
27
Different Initializations
28
Occlusion
29

• Thank you