Research On Siamese Network Object Tracking Algorithm Based On Weight Adaptive Updating And Graph Network

Visual object tracking is a hot research topic in computer vision.It has a wide range of applications in unmanned driving,video surveillance,human-computer interaction,intelligent transportation and other fields.Although the object tracking algorithm has been greatly developed,in the actual application environment,the object tracking algorithm still faces many challenges,such as object deformation,scale change,rotation,similar object interference,etc.,so the performance of the algorithm still needs to be further improved.Based on the Anchor-free Siamese network object tracking algorithm,the main research contents are summarized as follows :(1)In order to solve the problem that the Siamese network object tracking algorithm relies too much on the initial frame information and lacks the effective use of online data,this paper designs a object tracking algorithm based on weight adaptive update module and template filtering.Firstly,the classification branch feature map is processed by spatial attention and channel attention respectively,and the processed feature map is fused.Secondly,the fused feature map is passed through the filter module,and the filter parameters are updated by using the online tracking stage data to suppress the background noise in the tracking stage.Finally,the template screening module is used to compare the feature maps of the first frame template branch and the updated template branch.If the result is greater than the set threshold,the filtered template replaces the initial frame template to complete the subsequent object tracking task.The algorithm is tested on standard data sets such as OTB100,UAV123,VOT2018 and VOT2019.The experimental results prove the effectiveness of the algorithm.(2)The traditional Siamese network object tracking algorithm uses cross-correlation or deep cross-correlation to measure the similarity between template frame and detection frame,which makes the algorithm unable to effectively adapt to extreme object deformation.This paper designs a object tracking algorithm based on graph network and Io U perception.Firstly,based on the improved Res Net50,the(Normalization-based channel attention module)NCAM is introduced after each residual structure to construct a feature extraction network with channel adaptive adjustment.Secondly,based on the graph network,a new similarity calculation method between template frame and detection frame is designed.The similarity calculation of the graph network nodes of template features and detection features is carried out to obtain the feature response graph.Finally,in the classification and regression part,the Io U-aware classification loss function is used in the classification branch to establish the correlation between the classification branch and the regression branch.In the regression part,(Complete Io U)CIo U is used to calculate the regression loss in the offline training phase.The algorithm is tested on standard data sets such as OTB100,UAV123,VOT2018 and VOT2019.The experimental results prove the effectiveness of the algorithm.