Rotation-Aware Discriminative Scale Space Tracking

Although discriminative correlation filters (DCF) based methods have achieved desirable performance in visual tracking, most existing methods still fail to cope with significant target rotations in challenging scenarios. This paper proposes a rotation-aware DCF-based tracking method that efficiently estimates the target rotation and exploits the estimated rotation in translation and scale models. The proposed method reformulates the discriminative scale space tracker (DSST) to estimate target location and rotation, simultaneously. Then, the estimated rotation integrates into the scale estimation process, effectively. While the proposed method can estimate all target rotations, it uses only three target models in each frame to have desirable tracking speed. Experimental results demonstrate that the proposed method outperforms the average precision and success rate of the DSST up to 3.9% and 5.3% on 21 relevant video sequences, respectively.