Anti-occlusion target tracking algorithm based on correlation filter architecture

Abstract

To address the issue of tracking failure facing tracking algorithms due to object occlusion or deformation, an object tracking algorithm is proposed on the basis of positional joint filter bank and inter-frame prior information estimation. Based on the tracking framework of discriminant scale space tracking (DSST), A set of filters are introduced to track specific sub-regions of the tracking target to solve the tracking jitter problem of target due to partial occlusion under a single filter tracking; At the same time, the inter-frame transcendental information of video target is introduced to reckon location of target to solve tracking failure, which effectively addresses the problem of fully occlusion of target, and then different tracking strategies and update models strategy are selected according to the average peak correlation energy (APCE); Finally, in comparison with other tracking algorithms on dataset OTB2013,for example, benchmark tracker, the proposed algorithm improved the overlap success rate by 23.6%, and average accuracy by 31.7% under deformation and occlusion, which can effectively solve the issues of target occlusion and scale challenge and improve accuracy of target tracking.