Research On Single Object Tracking Based On Siamese Network

Computer vision is regarded as an extension of human visual perception,which can automatically obtain,analyze and understand useful information from digital images.As one of the successful application areas of deep learning,computer vision already has a large number of practical applications.Among them,visual object tracking based on video images is one of the key research directions in computer vision technology,which can provide new ideas and methods for advanced visual understanding problems such as motion analysis,event detection,trajectory prediction,and activity recognition.This article first systematically studies and summarizes the single object tracking algorithm based on deep learning and related filtering technology,and proposes two object tracking algorithms:In order to solve the problem of lack of large-scale training sets during the current mainstream tracker training,a self-supervised tracker that can use unlabeled video images for network training and learning is proposed.The tracker adopts the method of embedding the discriminative correlation filter into the twin network,and performs the offline training of the network through the loop consistency loss composed of forward tracking and backward tracking,and by adding a subspace attention mechanism and a channel attention mechanism to the feature extractor,the discriminability of the object feature is improved.The test results on the single object tracking data set show that the proposed tracker can extract effective object features through self-supervised learning,and can reach the tracking benchmark of the supervised tracker;Aiming at the problem that the existing tracker combined with related filtering technology has insufficient ability to express the object features and in the tracking process,the tracker has problems such as target drift caused by target template pollution,a object tracker that combines dynamic convolution and elastic update of target templates is proposed.Without increasing the width and depth of the network,the dynamic convolution structure improves the feature expression ability of the convolutional network for tracking objects.And by using the calculation of the tracking confidence,through the selection and update of the target template,while improving the tracking speed,it effectively copes with the complex background and target occlusion in the tracking problem.