Real time tracking via sparse representation

The L1 tracker gains robustness by casting tracking as a problem of sparse approximation in a particle filter framework. Unfortunately, the particle filter and ℓ norm minimization lead to a large amount of calculation as a result that the L1 tracker cannot achieve real-time tracking. The aim of this paper is to develop a new tracker which not only runs in real time but also has a better robustness than L1 tracker via sparse representation. In our proposed algorithm, candidate targets are sampled in the region of interest (ROI) to increase the tracking speed. Moreover, based on the block orthogonal matching pursuit (BOMP), a very fast solver is developed to solve the problem of ℓ norm minimization to improve tracking speed and accuracy. We conduct extensive experiment to validate and compare the performance of the BOMP algorithms against six popular ℓ-minimization solvers in different challenging sequences. We also implement great experiment to validate the high computational efficiency and tracking accuracy of our proposed tracker compare with four alternative state-of-the-art trackers in six challenging sequences.