Kernel correlation filters based on feature fusion for visual tracking

Abstract

Abstract This paper presents a CF-based method for robust visual tracking by multiple features fusion. HOG features and high-level CNN features are used to enhance tracking performance. First, different features are extracted from the target. Secondly, the extracted features are employed to learn the corresponding discriminant models, which are used to generate response maps respectively. Finally, an accurate response map can be obtained by fusing multiple response maps. In addition, we propose a model update strategy to reduce model drift under occlusion. The OTB benchmark is used to test the proposed method. Experimental results show that the proposed method can significantly improve tracking performance in complex situations especially in occlusion and fast motion. In addition, compared with several state-of-the-art methods, the proposed method shows competitive accuracy rate and success rate.