Target Tracking Algorithm Based on Multi-feature Adaptive Fusion in Complex Scenes

Traditional correlation filtering algorithms limit the tracking performance in complex scenes because they cannot fully utilize the features of the target. To solve this problem, a multi-feature adaptive fusion target tracking algorithm is proposed in this paper. First, the HOG feature, CN feature and Gray feature of the target are extracted respectively, and the CN feature and Gray feature are linearly splined into color feature by cat function to improve the discriminability of feature representation in the tracking process. Then, the maximum response values of HOG feature and color feature are calculated respectively, and the response values are normalized. Finally, the fusion weights of the features are determined according to the normalization results, and the adaptive fusion of the features is realized in the response layer. In order to ensure the feasibility of feature fusion, APCE value was introduced in this paper to compare the indicators before and after fusion, and experiments were conducted on OTB2015. The results show that the algorithm presented in this paper has significantly improved the success rate and accuracy, and has good robustness to deal with complex environments.