Probabilistic graphical models for visual tracking of objects
Vijay Badrinarayanan · OpenGrey (Institut de l'Information Scientifique et Technique) · 2009
This thesis puts forth graphical models for visual tracking in low and higher dimensional state spaces. For low dimensional tracking problems, such as object position tracking, a novel message switching/combination idea is introduced. Based on this concept and a new pseudo simulation viewpoint of point trackers a novel randomized feature point filter is developed. The message switching/combination ideas are then extended to construct multi-cue fusion based tracking with a set of simulation based filters. Employing pseudo simulation based point trackers and color based particle filters as their elementary filters these multi-cue fusion schemes track general objects in complex scenarios. Moving to higher dimensions, general multi-part tracking schemes are introduced. A network of local patch trackers are put to play in a stochastic simulation framework to track the position and other attributes of arbitrary objects. The difficult task of updating the process prior online is also performed under this simulation framework. Dealing with online update of the process prior enlarges the scope of application of the multi-part tracking model. A simple interactive multi-part tracking scheme is also discussed in this context. To the extent permitted by practicality, the contributions are evaluated quantitatively and/or qualitatively to convince the reader of their novelty, improvements and/or robustness. Detailed discussions of the highlights and drawbacks of the models are presented. Prospective extensions of the models based on empirical arguments and reflections on design are included.