Abstract:

A robust visual tracking algorithm that incorporates the probabilistic principal component analysis (PPCA) appearance model in a particle filter framework is proposed. To effectively model the large changes of scene illumination and object appearance, an online learning PPCA appearance model is used for acquiring subspacebased object representation, and a forgetting factor is also introduced in order to increase the importance of latest observations in the appearance model. Occlusions are handled by partitioning the object region into blocks considering their spatial adjacency. Experimental results on real complex situation demonstrate that the proposed algorithm tracks objects well no matter when they are under large pose variations, illumination changes, or severe occlusions.