| In recent years,with the rapid development of computer vision,self-driving,action analysis,intelligent transportation and intelligent robots and other fields,which have put forward higher demands for visual tracking.Single target tracking is a hot field in computer vision.After select the tracked object,it uses the context information of a video or image sequence to model the appearance and motion information of the tracked object,so as to predict the motion state and position of the tracked object.Generally,single target tracking algorithms can be divided into three categories: methods based on generative model,methods based on discriminative model and deep learning.In order to achieve better performance,deep neural networks like Res Net and others have been introduced into the field of visual tracking instead of Alex Net network,to obtain more representable features through large datasets training.However,there are great developments on the new object tracking algorithms,the introduction of RPN network and deep neural networks bring new problems,such as the mismatch between classification branch and regression branch of RPN and bad performance with Res Net.Tracking drift will occur when the tracking target is occluded,there is no effective solution for existing algorithms now.To solve above problems,this paper proposes a video single target tracking algorithm based on Siamese network,and the main contributions are as follows: this paper use the adjusted Res Net50 network as feature extraction,using deformable spatial aware module to fuse multilayer features,making full use of the shallow details information and deep semantic information,suppressed the influence of mismatch problem of RPN network;a large number of difficult occlusion samples were generated by Cutout during the target template branch acquisition in training for the target occlusion problem,which further improved the robustness of the algorithm.the self-adaption modules that weight sharing are added to two branches of Siamese network to extract the internal implicit correlation features about images data.Finally,it is designed as a software application to realize its application value.We compared the performance of classical algorithms,excellent tracking algorithms in recent years and the more accurate target tracking algorithm(ASiam RPN)on the VOT dataset in this paper,including accuracy,robustness and expected average overlap rate.Experimental results show that ASiam RPN proposed in this paper can effectively suppress occlusion and similar object interference,and the tracking accuracy is significantly improved to 60.9% on VOT2018,which is better than other trackers,and the speed is about 25 FPS. |
