Research On Deep Learning Object Tracking Method Based On Siamese Network

Object tracking is one of the basic tasks of computer vision.It has broad application prospects in the fields of autonomous driving,video surveillance,human-computer interaction,and augmented reality.Therefore,object tracking has attracted more and more attention and research.In recent years,the research on object tracking has made great progress,but the performance of related algorithms still needs to be improved.The tracking effects of target occlusion,deformation,background clutter,scale,illumination,blur,and rotation are not satisfactory.Object tracking only provides the appearance and position information of the first frame,and requires the tracker to track the subsequent frames continuously and stably.Therefore,how to improve the accuracy and robustness of the algorithm in complex tracking scenes based on limited information and the anti-jamming ability of similar interferences is still an important research direction in the field of object tracking,and the research of related theories and methods has very important significance and value.Based on the deep learning technology,this paper conducts research and exploration on the feature extraction method and template update method on the basis of the Siamese Network's object tracking algorithm.The main innovations are summarized as follows:(1)In the process of feature extraction,traditional convolution operations will gradually tend to focus on global information and ignore local information,which causes the problem of insufficient feature discrimination.This paper proposes a feature extraction strategy based on striped convolution,which splits the feature maps in specific dimensions,and re-splices them in the original order after convolution.This strategy helps the model learn the local information of the target,acquire features with higher semantics and more stability,and improve the robustness of the features.The experimental results show that compared with the traditional convolution operation,with the increase of the number of training iterations,the striped convolution can make the network pay more attention to the local information of the target in the process of feature extraction,obtain features with stronger discriminative ability,and improve the network to distinguish similar targets.(2)The method of fusing the tracking target template of each frame in Update Net requires the model to find the difference among many similar targets and update the template with the difference information.This not only increases the learning difficulty of the model,but also does not pay attention to the difference information between the template and the surrounding distractors.In this paper,a template updating method based on multi-frame gradient information is proposed.Firstly,the gradient is obtained by using the classification results predicted by the model and the GT simulated by Gaussian function,and the template is updated by making full use of the target feature change information and the difference information between the target and the surrounding distractors contained in the gradient,then the multi-frame gradient is fused by network learning.This method can adapt to the complex tracking scene,and compared with the fusion of each frame tracking target template method,this strategy can reduce the difficulty of learning.The accuracy of target tracking and the robustness of the algorithm are effectively improved.