Tracking by local collaborative representation

This paper proposes an object tracking method based on local collaborative representation. In this method, local image patches of an object are represented by collaborative vectors with an over-complete dictionary and a classifier is learned to discriminate the target object from the background. To deal with the changes of both the target object and the background during tracking, we update the dictionary and classifier with new observations obtained online. Compared to the recent tracking algorithms based on sparse representation, the collaborative representation method is also effective yet much more efficient. Furthermore, collaborative representation based on local image patches and the discrimination formulation the proposed algorithm employs can deal with complex environments better than the sparse representation-based tracking methods which use holistic object appearance within a generative framework. Experiments on lots of challenging video sequences with comparison to several state-of-the-art methods show the favorable performance of our algorithm.