Research On Vision Single Object Tracking Technology Based On Deep Learning

As a significant task in the field of computer vision,obiect tracking has a wide range of applications in the fields of Virtual reality,intelligent transportation and UAV.With the further development of machine learning technology,the object tracking technology has been greatly improved,but since the problems of object deformation,occlusion and background clutter,achieving accurate and robust tracking in complex scenes still faces great challenges.For purpose of enhancing the localization accuracy and robustness of the siamese network object tracking algorithm,the thesis studies from the following two aspects.(1)To solve the issues of object deformation and similarity interference in complicated views,based on the Siam RPN algorithm,an object tracking algorithm MFAW-Siam RPN based on siamese network feature fusion and adaptive weighting is designed.A multilayer feature fusion network based on the skip ASPP module is imported to extract features from search image and template image,and a feature adaptive weighting network is used to The adaptive weighting network is used to weight the feature map,which can selectively select the feature information of the image according to the importance of the information and screen out some interference information when the similar object interference exists.Finally,introducing the free-anchor classification regression module at the end of the network predicts the bounding box information of the object pixel by pixel to reduce the computational effort.The experimental tests demonstrate that compared with the Siam RPN algorithm,the improved algorithm MFAW-Siam RPN jumps he expected average overlap rate on the VOT2018 dataset by 2.3 percentage points,the success rate and tracking accuracy on the OTB2015 dataset jumped by 2.8 percentage points and 2.5 percentage points,respectively.The tracking performance is better,with better robustness in handling complex scenes,and effectively improves the tracking ability in similar object interference scenes.(2)To address the problem of frequent disappearance and relocation of tracked objects in long-time object tracking scenarios,a long-time object tracking algorithm based on template update and redetection(LTUSiam)is designed based on the MFAW-Siam RPN algorithm.Firstly,based on the basic tracker MFAW-Siam RPN,a three-stage cascaded gated cyclic unit is introduced to judge the object state and select a suitable time to update the template information adaptively using a template update network.Secondly,a template matching-based redetection algorithm is proposed to design a new similarity judgment method for rough localization of the global image,and then a precision localization module is used to classify and refine the bounding box of all candidate regions based on ROI features.Experimental tests show that,the improved algorithm LTUSiam achieves a frame rate of 28 frames per second on the VOT2018＿LT dataset,which achieves better results in real-time tracking,while achieving a performance of 0.644 on the F-score,better robustness in dealing with the object loss reproduction problem,and effective improvement in long-time tracking performance.