Contextual Adaptive Particle Filtering for Robust Real-time Non-Rigid Object Tracking

Particle Filtering algorithm for tracking the location of an object using a color distribution is one of the most used algorithm in many sub-field of visual tracking problem. However, the use of a color distribution for tracked object description is insufficient in practice. In this paper, we present an adaptive contextual particle filtering algorithm integrating multiple cues to non-rigid object tracking, designed to handle illumination variation, scale change and complex non-rigid motion. For this purpose, low-level contextual information computed through tiara lick texture features and color cues are combined into a model describing the appearance of the target. The likelihood of each cue is calculated and the algorithm rely on likelihood factorization as a product of the likelihoods of the cues. Moving object extraction is performed at each frame for initializing the filter and adapting the search space of each particle with the real dimension of the tracked target. Experimental results of applying this approach show improvement in tracking and robustness in recovering from very complex conditions.