Research And Application Of Visual Object Tracking Based On Siamese Network

With the boom of deep neural networks,the field of visual object tracking is growing rapidly,and it has been widely used in the field of unmanned aerial vehicles,intelligent surveillance and unmanned driving,etc.Object tracking algorithms based on Siamese networks have attracted widespread attention from scholars at home and abroad because of it achieves a better balance in tracking robustness and time efficiency.Therefore,how to improve the robustness of the Siamese network object tracking algorithm has become the main research goal of most researchers.The main research target of this paper is the object tracking algorithm based on Siamese network and its improvement.The specific research content is as follows:(1)Aiming at the problem that it is difficult to adapt to object positioning in the environment with similar semantic information interference,an object tracking algorithm that is based on residual channel attention and multi-level classification and regression is proposed.First,multiple residual channel attention modules are introduced on the target template branch to enhance the target foreground features of different depths and suppress background features in the target template,and then a multi-level classification and regression network is constructed.The classification and regression results of different depth features are weighted and fused using multiple trained weights,so as to obtain more information about shallow features and make the positioning more accurate.(2)Aiming at the problem of object positioning drift caused by large target deformation in video sequence,this paper proposes an algorithm for target tracking in Siamese network based on global and local feature matching.Firstly,In order to construct the pixel-level mapping relationship between the target template and the search area,a local feature matching structure is added on the basis of global feature matching structure,and then a graph residual attention module is established to obtain part-to-part pixel-level mapping information.In addition,for avoiding the local matching features containing too much background information,a selection mechanism of template-aware feature is introduced in the target template branch of the local matching structure to get the target foreground features.Finally,the classification and regression results of global matching features and local matching features are weighted and fused by using trained weights to obtain more information with pixel-level,thereby improving the robustness and accuracy of the algorithm.(3)Based on the algorithm combination proposed in this paper,an object tracking system in intelligent video surveillance is designed and implemented.The system can realize the continuous and stable tracking to interested objects,which is conducive to the monitoring administrator to monitor the behavior of abnormal objects,improve the protection class of the security department,and can effectively deal with emergencies,meet the requirement of social security protection for intelligent monitoring system,and realize the application value of this study.