Tracking in multimedia data via robust reweighted local multi-task sparse representation for transportation surveillance

It is of great importance in smart transportation surveillance to track object reliably from multimedia streaming data. Sparse representation based target tracking methods often suffer from tracking failure when target is under occlusions, pose changes or illumination changes conditions. In this paper, we propose a novel robust reweighted local multi-task sparse tracking algorithm. In the algorithm, local patches of all candidate targets are represented as a linear combination of the corresponding local patches from the template dictionary. Furthermore, in order to efficiently capture the frequently emerging outlier tasks, we decompose the sparse coefficient matrix to two collaborative matrices to make sure that the same type of particles share the same sparse structure. Observing that the edge of the candidate object contains background information, this paper gives a lower weight coefficient to the reconstruction error regularization located in the edge of the local patches than the middle local patches. Experimental evaluations on challenging sequences demonstrate the effectiveness, accuracy and robustness of our proposed algorithm in comparison with state-of-the-art algorithms.