Research On The Object Tracking Across Cameras

With the rapid development of computer vision,intelligent video surveillance system has attracted wide attention.As a key component of intelligent video surveillance system,object tracking system has high research value and application value.At present,the single camera object tracking system is mainly used.Because of the limited monitoring range of single camera,it has great limitations,such as unable to track the object continuously.The cross-camera object tracking system overcomes the inherent defects of single object tracking system by increasing the number of cameras,and achieves long-term continuous tracking of objects.Based on the background,the thesis studies the cross-camera object tracking technology.Firstly,the thesis studies the object tracking technology in single camera scene,and then studies the cross-camera object matching technology with shallow overlapping regions.The main work of this thesis is as follows:1.Aiming at the performance degradation of KCF algorithm in complex tracking scenarios,an improved algorithm is proposed in this thesis:(1)In order to solve the problem that a single HOG feature can not meet the needs of many tracking scenarios,the thesis proposes a method combining HOG feature and CN feature to describe the object;(2)In order to solve the problem that the algorithm is not suitable for object scale change,the thesis proposes a method of adaptively adjusting object scale size combined with scale filter;(3)In order to solve the model drift in the tracking process,the thesis proposes a tracking failure detection based on response peak,and adjusts the template update strategy adaptively.The improved algorithm is tested on VOT dataset.The experimental result shows that the overall tracking accuracy of the algorithm is improved by 12%,and the frame rate is increased by 13 frames per second.2.Aiming at the problem of cross-camera object matching with shallow overlapping regions,the thesis analyses a method of obtaining homography matrix accurately,and then uses the homography matrix to perform homography projection based on the object midline.(1)Since four pairs of feature points are the minimum requirement for calculating homography matrix,the thesis compares the accuracy of homography matrix obtained by using four pairs of points and eight pairs of points.(2)In order to reduce the computational complexity of the algorithm,the thesis abstracts the object as a linear model and uses homography projection to complete the object matching.The experimental result shows that the homography matrix calculated with eight pairs of poings has higher accuracy,and the object accuracy based on the object midline reaches 80%.