Combining multiple features and enforcing joint sparsity have proven to be beneficial for robust tracking. In this study, a novel stereo vision and two-stage sparse representation-based method is presented. First, the colouring information-based features are augmented with a depth view in the appearance modelling of a target object. Unreliable features are then dynamically removed for robust feature-level fusion in the first stage of sparse optimisation. Next, the low rank constraint is imposed onto the objective function, which facilitates a more robust representation of the ensemble of particles over the pruned views. Finally, the authors propose to detect occlusion via depth-based histogram analysis to guarantee the effectiveness of the template update. Experiments are performed on two large-scale benchmark datasets: KITTI and Princeton. Authorsa?? approach achieves state-of-the-art results in the aspect of robustness and accuracy.
展开▼