Target tracking based on multiple appearance model fusion with sparse representation

Abstract

This paper proposes a target tracking method based on multiple appearance model fusion with structured sparse representation in particle filter framework. In the proposed method, the global and local features of the target are sparsely represented respectively by global and local templates, the spatial structure information of the target in the local appearance model is characterized by summation of the sparsely coding coefficients of each local patch of one candidate target on local target templates that have the same position as that of the local patch. Meanwhile, the global and local templates are updated respectively based on sparse representation as well as the incremental principal component analysis along with block detection. The proposed template update method can reduce computational complexity and improve the update efficiency of the appearance model. Experimental results on video sequence dataset show that the proposed tracking method decreases the mean center error to 6.64 pixels, and improves the mean overlap rate to 69%. The proposed tracking method can track target robustly when existing similar object, partial occlusion, and illumination or pose variation, and rotation.