Correlation Filter Tracking Algorithm Based on Object Saliency Guidance

Abstract

When dealing with object tracking in complex scenes, such as occlusion, illumination variation, deformation, and low resolution, the traditional correlation filter (CF) has some shortcomings in the object representation model and the relocation.To optimize the filter model and improve the success rate of the tracking algorithm, an appearance representation model and a saliency-guided relocation model were proposed in this study.An adaptive weighted multifeature fusion model was designed based on the analysis of the internal correlation between different feature representation and CF responses.The weight coefficients were allocated adaptively according to the response value, and the final object position was obtained under the constraint of penalty term.With the first frame and the latest tracking result taken as the guidance, the saliency of the object was detected by using the updated multi-layer cellular automata method to achieve the purpose of object relocation after tracking drift or failure.Finally, the size of the object was predicted according to the strategy of size detection based on filter response.The accuracy of the proposed algorithm was verified by comparative experiments on benchmark data sets.Results show that the appearance representation model of the proposed algorithm is effective.The overlap success rate and distance precision rate of the proposed algorithm are 0.673 and 0.724, which are better than the results of other comparison algorithms.The proposed algorithm effectively solves the tracking drift or loss caused by complex scenes, such as occlusion, illumination variation, and low resolution.This study eliminates the problem of object recognition caused by environmental changes and provides references for the anomaly detection of real-time traffic.