Object Tracking Of Siamese Network Based On Sketch Structure Rematch And Feature Update

Object tracking technology has been widely used in many fields,such as intelligent video surveillance,human-computer interaction,and robot visual navigation.In 2016,Bertinetto et al.proposed the SiamFC algorithm based on the siamese network,which achieved good tracking accuracy while ensuring real-time tracking.SiamFC algorithm attracted wide attention from researchers.In practical applications,the scenes of object tracking are usually complicated,and the SiamFC algorithm has a lot of problems need to be solved,such as similar semantic interference,object deformation,and object occlusion.Therefore,this thesis makes the following research.(1)Aiming at the problem of multiple peaks in the response map caused by similar semantic interference in the SiamFC algorithm,which leads to tracking failure,a object tracking algorithm with siamese network based on maximum filtering multi-peak location and oneshot matching(SiamMPM)is proposed.The algorithm consists of two stages.The first stage is to obtain the image blocks corresponding to multiple peaks in the response map of the video frame through the proposed multi-peak positioning module based on maximum filtering.The second stage is to identify the object to be tracked from all image blocks in the first stage.This stage uses all the image blocks obtained in the first stage as the support set,the template image as the query set,and uses the one-shot learning method based on the matching network for re-matching.On the OTB2013 and OTB2015 datasets,we conduct ablation experiments and comparative experiments on the algorithm.Compared with the SiamFC algorithm,the accuracy of our algorithm is improved by 3.1% and 2.8% respectively,and it can effectively alleviate the similar semantic interference problem.(2)Aiming at the problem that the SiamMPM algorithm cannot effectively track the object when the object is deformed or occluded during the tracking process,a object tracking algorithm with siamese networks with feature update based on attention fusion mechanism(SiamUAF)is proposed based on the SiamMPM algorithm.SiamUAF algorithm uses the attention mechanism to fuse the features of the corresponding image of the video frame prediction boxes during the tracking process,and then uses the fused features to update the feature of the template,so that SiamUAF algorithm can accurately track the object when the object deformation is large or partially occluded.On the OTB2013 and OTB2015 datasets,we conduct ablation experiments and comparative experiments on the algorithm.Compared with the SiamMPM algorithm,the accuracy of our algorithm is improved by 1.9% and 2.1%respectively,and it can effectively alleviate the problem of large deformation or occlusion of the object.(3)In order to further improve the performance of the SiamMPM algorithm,by introducing the structure and position information of the tracking object in the sketch space,a object tracking algorithm(SiamFSF)is proposed,which combines sketch rematch and feature rematch.The first stage of the SiamFSF algorithm is the same as the first stage of the SiamMPM algorithm.In the second stage,firstly,the one-shot learning method based on the matching network is used to calculate the feature rematch scores of all image blocks acquired in the first stage.Then,the sketch rematch scores are calculated by the proposed sketch rematch module.Finally,through the proposed fusion strategy,the feature rematch scores and sketch rematch scores are merged to complete object tracking.The second stage of the algorithm uses not only the semantic information of feature but also the structure and position of the tracking object in the video frame.On the OTB2013 and OTB2015 datasets,we conduct ablation experiments and comparative experiments on the algorithm.Compared with the SiamMPM algorithm,the accuracy of our algorithm is improved by 1.1% and 1.3%respectively.