Object tracking with sparse representation and annealed particle filter

Recently, the L1 tracker is proposed for robust visual tracking. However, L1 tracker is still in traditional particle filter framework. As we know, particle filters suffer from some problems such as sample impoverishment. In this paper, we propose a new visual tracking algorithm, sparse representation based annealed particle filter, to further improve the performance of L1 tracker. As in L1 tracker, we find the tracking target at a new frame by sparsely representing each target candidate with both target and trivial templates. The sparsity is achieved by solving an (ell _{1}) -regularized least squares problem. The candidate with the largest likelihood is taken as the tracking target. But different from L1 tracker, instead of tracking objects in the common particle filter framework, we solve the sparse representation problem in an annealed particle filter (APF) framework. In the APF framework, the sampling covariance and annealing factors are incorporated into the tracking process. The annealing strategy can achieve “smart sampling” to avoid generating invalid particles corresponding to infeasible targets. Both qualitative and quantitative evaluations on challenging video sequences are implemented to demonstrate the favorable performance in comparison with several other state-of-the-art tracking schemes.