Research On Improved Agorithm Of Single Target Tracking Based On Deep Siamese Network

Single target tracking technology is one of the crucial technologies in computer vision tasks.With the rapid development of deep learning,single target tracking based on deep siamese network has gradually become the mainstream in the field of single target tracking.However,developing an accurate and real-time tracker still faces great challenges.When facing complex scenes,the tracker is difficult to track the target;Most trackers do not update the template information,which makes it difficult to track the occluded and greatly deformed targets;The correlation operation is local linear matching,which limits the modeling of complex nonlinear relationship between template branch and search area branch,and it is easy to lose useful information;In addition,the two branches calculate independently without information exchange,which limits the improvement of tracking accuracy.To solve the above problems,two tracking algorithms are proposed to deal with many challenges.(1)Aiming at the problem of difficult target tracking in complex scenes and lack of template updating,on the basis of deeper and wider siamese networks for real time visual tracking(Siam DW),a tracking algorithm named simple yet better methods to enhance siamese tracking(Siam SYB)is proposed.By integrating the attention mechanism module,the accurate characteristic response map is extracted and the network pays attention to the target;The template updating module is introduced to combine the historical frame,initial frame and prediction frame,so that the tracker does not have the phenomenon of tracking frame drift in the challenge of object occlusion and object deformation;The multi-stage model training strategy is adopted to train the accurate tracking model step by step.(2)Aiming at the problems of no information exchange between the two branches of siamese network,the disadvantages of traditional cross-correlation operation and insufficient utilization of multilayer features of convolutional neural network,and based on solving the problem in(1),a tracking algorithm named graph attention information fusion for siamese adaptive tracking(GIFT)is proposed.Among them,based on the attention mechanism,a siamese adaptive attention(SAA)module is proposed,which enables the two branches to exchange information and guide each other to adaptively extract tracking information by exchanging the channel attention weights of the two branches.SAA module indirectly updates template information;Based on the graph attention mechanism,a graph attention information fusion(GAIF)module is proposed to effectively combine the information of the two branches;The feature output of each layer of convolutional neural network is fused by hierarchical aggregation strategy,which makes full use of multi-layer feature information to further improve the tracking performance.