Temporal consistency object tracker with ranking mechanism

Visual tracking is important in the field of robotic controlling. However, developing a object tracker, which is robust in complex scenarios, is still a challenging work. In the paper, we propose a novel structural local sparse representation based residual error consistent ranking tracker. In the particle filter framework, candidate targets are linearly combined by a local sparse dictionary. By exploiting temporal consistency, the proposed algorithm develops a residual error consistency term to constraint the objective function of sparse representation. The alignment-pooling algorithm is used to obtain pooled features which contain similarity information of candidates. For further robustness, we develop residual error scores to evaluate the likelihood of candidates belonging to targets. For different natures of the residual error scores and pooled features, a ranking mechanism is proposed to fuse them. Furthermore, the dictionary updating scheme, which combines incremental subspace learning and sparse representation together, uses ranking results of predicted targets to decide which of the predicted targets are collected for dictionary updating. Finally, the proposed tracker performs favorably against 6 state-of-the-art trackers on 6 challenging video sequences.