Research On Visual Tracking Methods Based On Deep Reinforcement Learning

Visual Object tracking is one of the important research topics in the field of computer vision.Deep reinforcement learning theory has been applied in the field of visual tracking in recent years.In this paper,several deficiencies of the existing visual tracking methods are analyzed,including the lack of experimental comparison of the detection and recovery strategies for target loss,only bounding boxes perpendicular to the coordinate axis can be predicted,and the super-parameters need fine adjustment,not flexible enough,with corresponding experiments conducted or new improvement schemes based on deep reinforcement theory put forward.The experimental results show that the proposed tracker improves the robustness and accuracy of the reference trackers to a certain extent.The main research contents and contributions are as follows:(1)This paper compares and analyzes the detection and recovery mechanisms of traditional visual object tracking methods for tracking scenarios in which the target is lost.Combined with a series of experiments on TC128 dataset,the validity of detecting target loss by using correlation-based method,response-based APCE method and other methods is compared by introducing evaluation criteria of classification problem.The effectiveness of dense sampling,gaussisan distribution sampling,RPN network and other methods for target loss recovery are compared.And a comprehensive analysis of the above schemes is given.(2)The weakness for dealing with rotation of the target of traditional visual tracking methods are analyzed.This paper designs and implements a target bounding box rotation prediction method based on deep reinforcement learning.An actor prediction network is trained using PPO reinforcement learning algorithm,which takes samples from video sequences as input,and produces angle predictions for target bounding box.Through a series of performance tests on the VOT dataset,it is proved that the proposed method can predict the rotation angle of the target frame to some degree.(3)Modern target tracking methods always need to set super-parameters precisely in order to achieve the best perfromance.Based on the idea of deep reinforcement learning,a network structure is designed and implemented,and a deep-reinforcement-learning-based training method is designed.The experimental results on the OTB dataset show that the tracking robustness of the proposed tracker is higher than that of the reference tracker with fixed hyperparameters in the whole and difficult situations,which validates the effectiveness of the proposed method.