Research On Visual Tracking Based On Deep Siamese Network

Visual target tracking technology is an important research direction of artificial intelligence.At present,it has important research value and significance.The method based on siamese network has achieved great success in the field of target tracking,but it still fails in the face of complex tracking scenes.How to build an efficient and robust target tracking system is still a great challenge.Based on the theory of deep siamese network framework,this paper carries out research work on strengthening the integration of network branch input,network structure and hierarchical features,and combining with deep residual network.The specific research work is as follows:(1)Propose a siamese network tracking algorithm based on attention mechanism,which enhances the discrimination ability of the network model,realizes online learning of target appearance changes and background suppression,and obtains robust tracking results.By adding the result obtained by tracking the previous frame into the template branch and the search branch as the correction unit,To make up for the shortage of the network in dealing with the appearance changes of the target,and to realize the feature fusion between different frames by adding the spatial attention module and the channel attention module to the Siamese network,thus learning the target deformation and background suppression online,and further improving the feature expression ability of the model.(2)Taking ResNet network as the basic network,the internal residual unit is introduced to modify the network structure to make it more suitable for target tracking tasks.Deep features contain rich semantic information,It is robust to the obvious appearance changes of the target,so the feature pyramid is introduced to fuse the upper and lower features of the siamese network to obtain feature maps rich in spatial geometric information and semantic information,and multi-dimensional feature expressions are generated for feature expression structures with different dimensions of the same size,thus obtaining more robust tracking results.(3)Contrast experiments with other algorithms in terms of tracking accuracy and success rate on the tracking data set,and contrast experiments are carried out under different influencing factors such as occlusion interference and deformation interference.The experimental results show that the improved algorithm proposed in this paper can effectively improve the tracking accuracy,and also provides an improved reference for solving the problems of similar background interference and target appearance changes that often occur in target tracking algorithms.