Learning correlation filter with fused feature and reliable response for real-time tracking

Object tracking is a key component of machine vision system and getting much attention in different walk of life. Recently, correlation filters have been successfully applied to visual tracking. However, how to design effective features and deal with model drifts remain open issues for online tracking. This paper tackles these challenges by proposing a real-time correlation tracking algorithm (RCT) based on two ideas. First, we propose a method to fuse features to more naturally describe the gradient and color information of the tracked object, and introduce the fused feature into a background-aware correlation filter to obtain the response map. Second, we present a novel strategy to significantly reduce noise in the response map and therefore ease the problem of model drift. Systematic comparative evaluations performed over multiple tracking benchmarks demonstrate the efficacy of the proposed approach. The results show that the proposed RCT significantly improves the performance compared to the baseline tracker while still maintaining a real-time tracking speed of 26 fps in MATLAB implementation.