Research On Target Tracking Based On Improved Siamese Network

Target Tracking is a task that continuously tracks the target of the video sequence.Only one frame image given by the video is used to track the target in the video subsequent frame by the algorithm.Target tracking has always been a research hotspot in the field of artificial intelligence.At present,target tracking has been used in video surveillance,autonomous driving and military UAV.With the wide application of deep-learning technology in target tracking,a large number of tracking algorithms with excellent performance have emerged,the most representative of which is the fully convolutional Siamese network target tracking algorithm Siam FC.The early Siamese network algorithms mostly use shallow the feature extraction backbone network of the layer still has the problem of poor tracking effect in the face of complex scenes such as target occlusion,scale scaling,and drastic changes in light intensity in real life.In addition,the tracking speed of the algorithm must also be considered.Therefore,in view of the above problems,this thesis is based on the Siamese network to study the target tracking,and the work is as follows:(1)A tracking algorithm Deep Siam based on multi-layer feature weighted fusion is proposed: Firstly,the algorithm abandons the alexnet network and adopts the improved Res Net-50 as the backbone network to extract more sufficient and complex target features by using the deep neural network.Secondly,the backbone network is weighted and fused by layers to obtain low-level contour features and high-level semantic features,and enhance the feature map’s ability to represent objects.Experiments show that the proposed method achieves better tracking performance on the target tracking test dataset and effectively improves the robustness of the algorithm in dealing with obscured targets.(2)An enhanced hybrid attention-based tracking algorithm ESASiam is proposed: combining the focusing ability of channel attention and spatial self-attention in deep convolutional networks to highlight regions of interest in images.Channel attention allows the network to capture the feature connections between different channels and highlight the channel features that are conducive to expressing the target.Spatial self-attention allows the network to establish a correlation between any two different pixel locations,helping to locate the target.In addition,a template-search collaborative attention module is designed to update template features implicitly,and an hourglass network is used to perform multi-scale information interaction between template features and search features extracted by the backbone network,to improve the robustness of the algorithm to handle changes in the target scale.(3)A tracking algorithm STASiam based on combining spatio-temporal context information is proposed: the temporal information in the video is aggregated by constructing a Laplacian template feature set,and forward propagated to the search area features through the cross-attention module.For the search area features,the template feature set is regarded as a mask that aggregates the features of different template frames,which is beneficial to make full use of the context information between video sequence frames,so as to improve the performance of the algorithm and solve the problem that Siam FC template cannot be updated.