Robust Tracking Based on Multi-feature Fusion

Abstract

In target tracking, only the single feature description target and the traditional model updating method are difficult to adapt to the occlusion, deformation and various complex scene changes of the target. To solve this problem, a correlation filter tracking algorithm based on multi-feature fusion and selective update model is proposed. Firstly, the filter model is trained by direction gradient histogram and color feature respectively, and two features are fused according to the peak sidelobe ratio of the different characteristic response maps in the detection stage. The target location of the target is judged to be blocked by the peak sidelobe ratio of the final target position response map of each frame, and the model does not update when the target is blocked. The current model continues to be tracked in the next frame. A comparison experiment is made between 12 challenging video sequences and multiple forward moving target tracking algorithms. The results show that the average center position error is reduced by 25.12 pixels compared with the suboptimal Color Names, CN, and the average tracking accuracy is increased by 29.31%. The experimental results show that the algorithm can track the target steadily and accurately in the case of scale changes, occlusion, and illumination changes.