Adaptive sparse vector tracking via online Bayesian learning

Abstract

In order to construct a flexible representation for robust and efficient tracking, a novel real-time tracking method based on online learning is proposed in this paper. Under Bayesian framework, RVM is used to learn the log-likelihood ratio of the statistics of the interested object region to those of the nearby backgrounds. Then, the online selected sparse vectors by RVM are integrated to construct an adaptive representation of the tracked object. Meanwhile, the trained RVM classifier is embedded into particle filtering for tracking. To avoid distraction by the particles in background region, the extreme outlier model is incorporated to describe the posterior probability distribution of all particles. Subsequently, mean-shift clustering and EM algorithm are combined to estimate the posterior state of the tracked object. Experimental results over real-world sequences have shown that the proposed method can efficiently and effectively handle drastic illumination variation, partial occlusion and rapid changes in viewpoint, pose and scale. © Springer-Verlag Berlin Heidelberg 2006.