Research On Visual Tracking Based On Deep Siamese Network

Target tracking,as basic research of artificial intelligence,has important applications in many fields.At present,the methods based on correlation filters and deep learning have achieved great success in the field of target tracking,but building a high-precision and robust target tracking system is still a huge challenge.Based on the theory of deep siamese network,this paper focuses on template selection,feature extraction and layering features in the network,and carries out research work with relevant filtering algorithms.The specific research work is summarized as follows:(1)A target tracking method combining the first frame and the previous frame template through the attention mechanism is proposed to obtain more robust tracking results under complex background.By combining the tracking result of the previous frame and the first frame as a template frame,the spatial attention module aggregates the feature maps of the two frames,and the channel attention module redistributes the weight of each channel to obtain the final template feature.And use deeper feature extraction network to extract features,realize large receptive field by stacking small convolution kernels,not only reduce the number of neural network training parameters,but also increase the depth of the network,and obtain more accurate target features.Experiments were conducted on public datasets,and the experimental results show an improvement in the robustness of the proposed algorithm.(2)A target tracking method combining background-aware correlation filter and hierarchical response map weighted fusion is proposed to obtain more accurate tracking results in complex background.The context-aware correlation filter is embedded into the deep siamese network.The quality of further feature extraction of the relevant filter layer is improved by introducing the context information of the target and directly merging into the learning of the filter template.Due to the existence of the relevant filter layer,the siamese network can obtain good tracking performance only by using shallow features,but the deep features contain rich semantic information and are robust to the significant appearance changes of the target.To this end,this paper proposes a weighted addition of the response graph generated by the shallow and deep convolutional features to obtain a more accurate representation of the target.The experimental results show that the proposed algorithm can get more accurate tracking results than the benchmark algorithm.