Research On Video Object Tracking Algorithm Combining Template Updating And Anti-Interference Module

Video object tracking is a research hotspot in the field of computer vision,which is widely used in civil and military fields such as automatic driving,human-computer interaction and UAV monitoring.In recent years,the object tracking algorithm based on Siamese network has achieved good tracking accuracy while ensuring the real-time performance of the algorithm,and has become the mainstream algorithm in the field of object tracking.However,in the complex tracking environment such as camera motion and background clutter,this algorithm still faces many problems to be solved.Firstly,in order to prevent the introduction of noise,most algorithms based on Siamese networks abandon the template updating operation,which makes it difficult for the tracker to cope with the changes in the appearance of the object and fails to give full play to the best performance.Secondly,the problem of analogue interference and occlusion affects the foreground feature information of the object,which causes serious interference to the tracker based on feature matching and even leads to tracking drift phenomenon.In order to solve the above problems,a large number of existing Siamese network object tracking algorithms are studied and analyzed in this paper,and two kinds of reliable video object tracking algorithms with template updating and anti-interference capabilities are proposed,which significantly improve the robustness and stability of the tracker.The specific research content and innovation of this paper are as follows:(1)A multi-template fusion object tracking algorithm based on GAT is proposed.Aiming at the problem of template updating in the process of object tracking,this paper proposes a multi-template fusion object tracking algorithm based on graph attention network through a lot of research and analysis of the existing template updating mechanism.Firstly,in order to effectively screen out reliable templates for template updating in the actual tracking process,a new two-stage template updating threshold judgment mechanism was proposed,and the Pearson Correlation Coefficient was introduced to supplement the existing threshold discrimination ability.Secondly,the feature embedding module based on graph attention network propagates the features of the initial template to each reliable template,suppressing the background noise beyond the object and improving the characterization ability of the template.Finally,the reliable template is embedded into the same feature space through the multi-template fusion module based on 3D convolution,and the latest template containing object space-time information is obtained for follow-up tracking.The comparative experimental results on multiple mainstream common data sets and the actual tracking effect fully verify that the proposed algorithm can update the template in time during the tracking process,significantly improve the robustness and stability of the tracker,and effectively alleviate the occurrence of tracking drift phenomenon.(2)A video object tracking algorithm based on interference aware and anti-interference modules is proposed.Aiming at the problem of interference and occlusion of similar objects encountered in the process of object tracking,based on the analysis and understanding of a large number of antiocclusion tracking algorithms,this paper proposes a video object tracking algorithm that integrates interference aware and anti-interference modules.Firstly,in order to enhance the antiinterference ability of the tracker,the Repulsion Loss is introduced into the regression branch to optimize the loss function and training process.Secondly,in order to enhance the ability of the tracker to perceive jamming objects,a new index reflecting the degree of proximity between the tracking object and the surrounding jamming objects,Positive Sample Prominence Rate,was designed,and the proposed index was combined with the existing threshold to efficiently screen jamming and occlusion frames according to the response map.Then,in order to enhance the antiocclusion capability of the tracker,a feature learning module based on adaptive mask was proposed to establish a partial level correspondence between the object and the template,and significantly reduce the dependence of the tracker on the overall template.Finally,the gragh attention network is used to enhance the object foreground features,and the high quality template is aggregated by the weight adaptive module based on channel attention,which further improves the robustness and stability of the algorithm.The algorithm proposed in this paper has achieved excellent tracking performance on mainstream public data sets such as OTB2015,GOT-10k and VOT2016.The test results of OTB2015 data set reached 0.700 tracking accuracy,which increased 41 basis points compared with the benchmark algorithm.