A rotation adaptive correlation filter for robust tracking

Tracking-by-detection methods have shown to provide excellent tracking performance, of which the approach using circulant structure with kernel (CSK) really stands out. The intrinsic idea of exploiting the circulant structure to model the procession of dense samplings, and applying the element-wise multiplication in Fourier domain, provides competitive performance and high possessing speed. However, the original method is limited only to estimate the target translation which may fail when object rotates and size varies at a large scale. This paper presents a strategy to enhance the tracker´s robustness against significant object rotations and scale variations. To tackle the problem of in-plane rotation, a multi-orientation CSK model is proposed, and a separate scale filter is used to handle scale estimation. Experimental results show that the proposed method outperforms the original KCF tracker by 7.04% with the same feature and 9.07% with feature fusion on Visual Object Tracking 2014 (VOT 2014) dataset.