Research On Visual Tracking Algorithm Based On Siamese Network

Visual tracking is an important research direction in the field of computer vision.It has been widely used in military,medical,robotic,intelligent transportation and so on.In recent years,deep learning technology has achieved great breakthroughs.The powerful learning ability of deep neural networks provides a new research idea for visual tracking,which makes the tracker have higher accuracy.The visual tracking algorithms based on deep learning has important theoretical and practical significance.The depth learning-based tracker has achieved great success and greatly improved the performance of the tracker.However,since the deep neural network involves complex and large calculations,the real-time performance of the tracker is difficult to guarantee.This paper is based on the Fully-Convolutional Siamese Network(Siam-FC),which consists of two identical sub-network branches.It has simple structure,outstanding performance and real-time tracking potential.However,the Siam-FC algorithm cannot handle the noise situation in the target template.Specifically,in the noise environment,when the feature map of the target template and the search region feature map are fully convoluted,there may be multiple maximum values in the obtained response graph,which makes the tracker extremely unstable.By designing two novel network architectures,this paper strengthens the response of the target region and suppresses the response of the noise region,thereby improving the tracking performance of the visual tracking algorithm in complex visual scenarios:1.Siamese of Cross Layer Contrastive Loss(Siam-CC).The Siam-CC network explores the feature properties of different depth network layers,and performs fully-convolutional operations on different layer features and specific layer features,so as to explore the features of different depth search regions by using the most suitable features of the target region.In order to take advantage of the deeper features of the search area,Siam-CC designed an asymmetric siamese network architecture to add a new network layer at the end of the branch of the search area.In addition,to enhance the discrimination between valid candidate frames and invalid candidate frames,Siam-CC proposes an adaptive contrastive loss function.The experimental results show that compared with the original Siam-FC,the tracking result of Siam-CC on the benchmark has been greatly improved.2.Multi-Domain Residual Siamese Network(MDR-Siam).The MDR-Siam network uses the residual block to filter the interference response points in the response graph,and improves the structure of the residual block by the group normalization method,so that the filtering effect is more suitable for the visual tracking task.On this basis,MDR-Siam uses a multi-domain approach to design the residual block as multiple independent domains,so that a particular domain is only responsible for processing a particular video.At the same time,the target template adaptively filters the different background noises,adaptively weakens the background noise of the target template according to the aspect ratio of the target in the template,and reduces the interference in the matching process.The experimental results on OTB-50 show that MDR-Siam has a certain improvement in tracking accuracy than Siam-FC.