Research On Bounding-Box-Level And Pixel-Level Single Object Tracking Methods Based On Siamese Network

With the development of artificial intelligence and multimedia,computer vision has been more and more widely used in various fields.As one of the important research directions in the field of computer vision,visual object tracking has been widely concerned by domestic and foreign scholars for many years.It has been widely used in the fields of intelligent video surveillance,human-computer interaction,medical image,unmanned driving,UAV and so on.In recent years,with the development of deep learning,a large number of algorithms based on deep learning gradually occupy the dominant position in the tracking field.Among them,the tracking algorithm based on Siamese network has attracted extensive attention because of its good balance in speed and performance.However,due to the lack of effective template update mechanism,its performance is far from that of tracking by detection algorithm.In addition,with the continuous development of the tracking field,the tracking based on the pixel-level target representation began to appear in the Siamese network algorithm.The accurate segmentation mask of the target object has brought greater challenges to the tracking algorithm.This thesis proposes two solutions based on the above two problems of Siamese network.1.A bounding-box-level object tracking algorithm based on dual template update mechanism.Firstly,the algorithm first starts with how to effectively use the tracked template,and uses the similar transformation of frequency domain between target templates to capture the changes in the appearance of the target to achieve template update.In order to meet the real-time requirements of the tracker,the similarity transformation matrix is established in the frequency domain to speed up the transformation.Secondly,in order to make effective use of the target templates of historical frames,we use the target templates extracted from these historical templates as the optimization target,and further update the target template.The experimental results show that the two template update mechanisms proposed in this thesis are better than some state-of-the-art tracking algorithms,which verifies that the proposed update mechanism can effectively deal with the changes of target appearance and improve the robustness of the tracker.2.Edge aware pixel-level object tracking.The algorithm starts with the edge information and uses the complementarity between the edge information and the target information to propose an edge aware tracking network,which includes a segmentation network branch to improve the edge segmentation accuracy of the target object.Firstly,in order to increase the robustness of the tracker,the cross-correlation features of the classification branches are fused with the high-level features in the backbone network,and the target edge and target segmentation mask are used for supervision at the same time to obtain the rough contour features of target.Secondly,in order to obtain more fine edge information,the rough contour feature of target is guided to the low-level feature in the backbone network to generate the fine edge features of target.Finally,in order to make use of the complementarity between target information and edge information,the fine edge features of target are fused with the target features of different layers to generate the target segmentation mask.Experimental results show that the performance of the edge aware tracking network proposed in this thesis is better than some state-of-the-art tracking algorithms,and can effectively improve the accuracy of target segmentation.