Research On Object Tracking Algorithm Based On Siamese Network

As one of the important and challenging tasks in computer vision,object tracking has important application value in object strike,medical diagnosis,virtual reality and other fields.In recent years,with the introduction of the Siamese network idea,object tracking has ushered in unprecedented development.However,when the object appears in a complex scene,the tracking performance of this type of algorithm is still not ideal.This paper mainly analyzes and studies the problem of poor performance of existing twin network tracking algorithms in complex scenarios such as similar interference,object deformation,and scale changes,and proposes two new tracking algorithms.The main results are as follows:(1)Aiming at the problems of weak feature extraction ability and unused template updating of Siam FC algorithm,a deep siamese network tracking algorithm based on template updating and residual attention is proposed.The algorithm replaces Alex Net with a deep Res Net-34 with an improved structure to extract more semantic features.At the same time,the ECA module is embedded in each residual block of Res Net-34.By considering the relationship between the adjacent channels of the convolution feature,an attention weight for the channel is learned,which increases the sensitivity of the network to the channel feature,thereby improving the performance of the network.In addition,an efficient template update strategy is introduced for template update,which is beneficial to deal with the impact of object appearance changes.The experimental results show that,compared with the benchmark algorithm Siam FC,this algorithm has better tracking performance in complex scenes.(2)In order to better cope with the scale change problem,an anchor-free siamese network tracking algorithm based on multi-feature fusion is proposed.The algorithm predicts the object frame in a pixel-by-pixel manner.It not only avoids the hyperparameter settings and complicated calculations related to the anchor,but also improves the adaptability to scale changes.In addition,the output features of the last three layers of the backbone network are selected to be adaptively fused,and the fused features provide more semantic information and detailed information for the subsequent prediction network,which further improves the tracking performance of the algorithm.The experimental results show that,compared with the benchmark algorithm Siam RPN,this algorithm achieves better tracking performance and can better cope with the scale change problem.