Adaptive Visual Tracking Based on Key Frame Selection and Reinforcement Learning

Abstract

When the appearance of a target changes dramatically, the traditional tracking methods can no longer be used for target tracking task. This paper proposes an adaptive visual tracking algorithm based on key frame selection and reinforcement learning (RL) to solve this problem. First of all, the probability value of the RL network output is analyzed, and the predicted value of output is normalized. At the model update stage of each frame, the probability value corresponding to the optimal behavior is judged. If it conforms to the preset rules, the last frame of the current frame is set as the key frame, and the network model is fine-tuned by using the key frame. The proposed algorithm is only fine-tuned in key frames to obtain multiple fixed prediction models. The experiment is conducted on 100 video sequences of the Object Tracking Benchmark to verify the effectiveness of key frame selection strategy. Compared with the original reinforcement learning based tracking algorithm, the tracking accuracy and the success rate of the proposed algorithm are improved respectively.