Improving the Speed and Robustness of Partial Occlusion Invariant Sparsity-based Algorithm of Visual Tracking with Coarse and Fine Appearance Modeling

The main goal of this paper is to find a partial occlusion robust tracking algorithm. In addition, this algorithm should have superior performance against other tracking challenges and should have real time operation. In according to these reasons, we propose a two stage tracker based on coarse and fine sparse representation. At first, target object appearance modeled by a dictionary consisting of PCA basis vectors and trivial templates. We apply APG_L1 solving method for this modeling. We decrease number of trivial templates to decrease computation load such that we consider one trivial template for every a by a pixels. Then, best candidates in previous stage are evaluated by PCA_L1 method in order to determine the last candidate. We compare this tracker with 5 different popular algorithms and 7 new sparse trackers using 12 datasets. We conclude that proposed tracker with 14.2 frame per second, 10.2 average pixel center error, 0.7475 average overlapped rate, has a better performance in comparison with other trackers. And therefore simulations show that proposed tracker has a better performance in both accuracy and speed in comparison to other algorithms.