Robust visual tracking with occlusion detection using compressive sensing

In this paper, tracking problem is considered as a sparse approximation of target by templates created during video process. In addition, some trivial templates are used to avoid the effects of noise and illumination changes. Each candidate is sparsely represented by the template set. This goal is achieved by solving an l₁- regularized least-square equation. To find tracking result, a candidate with the minimum reconstruction error was adopted. Then, tracking was continued in particle filter framework. Two ideas were used to improve the algorithm performance. Firstly, the dictionary set was adaptively updated according to appearance changes. Secondly, using the area around the target, occlusion was diagnosed and subsequently the template set was updated. This technique prevented the occluded part of the target getting into the template set. The proposed approach shows a better performance than other previous tracker against full occlusion problem.