Object Tracking with Shape and Motion Dynamics

1
Object Tracking with Shape and Motion Dynamics

• Sezen ERDEM
• Advisor Assoc.Prof.Dr. Sibel TARI

December - 2005
2
Outline

• Motivation
• Current Studies
• Our Approach
• Experimental Results
• Conclusion
• Future Works
• Questions

3
Motivation

• Detecting motion and monitoring moving objects in
  video streams is an important survalliance
  problem.
• There exist many problems
• Background clutter may mimic parts of foreground
  features
• Observation may contain noisy data

4
Motivation

• Solve the problems by using
• Object shape and motion dynamics.
• Probabilistic models of object shape and motion
  to analyse the video-stream

5
Motivation

• Our aim is to develop a tracking algorithm which
• Uses shape information
• Learns motion
• Uses motion dynamics to track target object

6
Current Studies

• Condensation Algorithm
• (Isard Blake)
• Activity Recognition Using the Dynamics of the
  Configuration of Interacting Objects
• (Vaswani,Chowdhury,Chelappa)

7
Our Approach

• Obtain Mean Shape of Object Using Statistical
  Shape Theory
• (Kendall Shape Model)
• Learn Motion Dynamics
• (Euclidean Motion)
• Use Kalman Filter to Track Object

8
Kendall Shape Theory

• All the geometric information that remains when
  the location, scale and rotational effects are
  filtered out from an object.

9
Kendall Shape Theory

• Configuration Matrix
• Configuration Space
• Position of Configuration Matrix
• Size of Configuration Matrix
• Pre-Shape Space
• Shape Space

10
Kendall Shape Theory

• Translation Normalization
• In order to make the shape invariant to
  translation, center the configuration matrix of
  the shape

11
Kendall Shape Theory

• Scale Normalization
• Normalize the configuration matrix by its
  Euclidean norm

12
Kendall Shape Theory

• Distance Between Shapes
• Generalized Procrustes Analysis (GPA) is used for
  measuring the distance between the shapes.

13
Obtaining Mean Shape

• Choose the first shape as an estimate of the mean
  shape.
• Align all the remaining shapes to the mean shape.
  
• Re-calculate the estimate of the mean from the
  aligned shapes
• If the mean estimate has changed return to step
  2.

14
Learning Motion

• Initially, a hand-built model is used in a
  tracker to follow a training sequence which must
  be not too hard to track
• Once a new dynamical model has been learned, it
  can be used to build a more competent tracker
• Generally two or three cycles suffice to learn an
  effective dynamical model

15
Learning Motion
Iterative Learning of Dynamics
16
Tracking with Kalman Filter

• After obtaining motion dynamics, use Kalman
  Filter for tracking.
• Kalman Filter
• Predict
• Measure
• Update

17
Experimental Results

• In our experiments we used
• a video segment with 125 frames.
• 25 pictures for learning mean shape
• 15 pictures for learning motion

18
Experimental Results
Result Video
Original Video
19
Conclusion

• Tracking approach in video streams which makes
  use of prior knowledge of shape and motion data
  of the tracked object
• Motion is represented by the configuration of the
  objects and its deformation over time

20
References

• Activity Recognition Using the Dynamics of the
  Configuration of Interacting Objects
• Vasvani,Chowdhury,Chellappa
• Articulation Priors A Feedback Mechanism for
  Object Recognition
• Erdem A.,Erdem E.
• Active Contours
• Andrew Blake Michael Isard
• Shape and Shape Theory
• D.G. Kendall, D. Barden, T.K. Carne, and H. Le
• Statistical Shape Analysis
• I.L. Dryden and K.V. Mardia
• Tools for Landmark Data Geometry and Biology
• Bookstein, F.L.

21
Questions