Researches On Object Tracking In Surveillance Video With Deep Reinforcement Learning

Currently,video surveillance is widely used in industrial and daily life.The explosive growth of video data brings unprecedented challenges to processing and analysis of such data.The traditional manual analysis of surveillance video data is unable to meet the needs of massive surveillance video processing because of its high cost,low efficiency and poor stability.Therefore,there is an urgent need to develop intelligent analysis technology.Object tracking is one of the most important topics of surveillance video analysis technology.It has important application prospects in the fields of intelligent transportation,intelligent healthcare,public safety and so on.Object tracking aims to locate certain objects in each frame.The key problems include the appearance description of the object,the objects association between adjacent frames and the computation efficiency to achieve real-time tracking.The tracking method based on deep reinforcement learning(DRL)combines the featureextracting ability of deep learning(DL)with the decision-making ability of reinforcement learning(RL)to provide a new framework for extracting more abstract and high-dimensional features and improving the tracking efficiency by using the relationship between frames,so that the method can realize real-time in an end-to-end way.In this paper,a single-object tracking method is proposed to solve the occlusion problem based on the theory of DRL,and then a multi-object tracking method is constructed by adding quadratic correlation algorithm to the single-object tracking method.The main contributions of this thesis are as follows:1)To improve the real-time performance of object tracking,an object prediction algorithm is improved and used to provide a more accurate initial location for object.The position of the current target based on previous tracking results is predicted to get a more accurate initial state of the object,which can significantly reduce the number of steps in the subsequent procedure of accurate object location.2)To solve the occlusion problem in single-object tracking,a tracking method based DRL is proposed with an occlusion frame detection branch to adaptively improve the tracking results in case of occlusion.The partial or fully occluded objects are accurately located by weighting the prediction of improved particle filter and the tracking result of current frame.3)To solve the multi-object mutual occlusion problem,a multi-object tracking method based on quadratic correlation is proposed to improve the tracking accuracy in the case of occlusion.When the object appears and disappears in the scene or occurs occlusion,multiple objects are matched twice to effectively associate with the corresponding historical object for improving the accuracy of multi-object tracking.