Research On Object Tracking Method Based On Deep Learning

Object tracking is one of the important research directions in computer vision.Object tracking is applied in a wide range of fields,including People’s Daily life,industrial production,national defense and military.The single object tracking algorithms based on Siamese network are simple in structure and excellent in performance.Such algorithms show certain advantages and have become the core direction of single object tracking field.However,trackers still face some challenges,such as occlusion,similar background interference and deformation,which lead to limitations in the accuracy and robustness of some trackers.In addition,many existing tracking algorithms are based on anchors,which have some problems such as excessive hyperparameters and poor generalization ability.To solve the above problems,based on Siamese network,this paper respectively proposed Siamese network based on spatial context module for visual tracking(SiamSCM)and Siamese network based on anchor-free region proposal network for visual tracking(SiamAF-RPN).The main work and innovation points of this paper are as follows:(1)To improve the accuracy and robustness of the algorithm,the spatial context module is introduced to obtain the spatial context information of the global view.Firstly,this paper analyzes the problems of baseline algorithm SiamRPN++.That is,the network does not fully consider the spatial dependence between long-distance pixels,resulting in the valuable spatial context information is not fully utilized,and the feature discrimination is weak,resulting in the accuracy and robustness of the tracker is insufficient.To address the problem,the spatial context module is introduced in this paper.Based on the idea of attention mechanism,this module enhances the discriminant ability of features by modeling global spatial context information,so as to improve tracking accuracy and robustness.In this paper,several datasets are evaluated,and good results are obtained,which shows the effectiveness of the proposed algorithm(SiamSCM).Compared with the baseline algorithm,the success rate and precision rate of SiamSCM on OTB100 dataset are improved by about 6.41%and about 3.21%,respectively,and the accuracy and average expected overlap rate on VOT2018 dataset are improved by 6.00%and about 7.20%,respectively.Compared to some good classical algorithms,SiamSCM has achieved excellent performance on GOT-10k and VOT2016 dataset.In addition,the robustness of SiamSCM is improved in a variety of complex scenarios such as occlusion,fast motion,and deformation.(2)In order to effectively solve the problems of excessive hyperparameters in anchor-based tracking algorithms,this paper introduces an anchor-free mechanism and proposes an anchor-free Siamese tracking algorithm(SiamAF-RPN).The algorithm no longer relies on anchors and achieves pixel-by-pixel classification and regression,which has lower model complexity and more stable tracking performance.Meanwhile,because of the unreliable bounding boxes that tend to be generated by edge pixels,this paper adds a central constraint branch to the head network of the tracker to penalize edge pixels.In addition,in order to make the algorithm adapt to the multi-scale variation of the object,this paper designs a feature combination network based on dilated convolution,which improves the performance of the tracker by constructing dilated convolutions with different dilated rates,thus acquiring features in different scale regions and improving the characterization ability of the features.Using SiamSCM as the baseline algorithm,SiamAF-RPN conducted ablation experiments on OTB100 and VOT2018 dataset to verify the validity of each module in SiarnAF-RPN.In addition,SiamAFRPN has achieved satisfactory results in comparison with some excellent classical algorithms on GOT-10k,UAV123 and VOT2016 dataset.Among them,on UAV123,SiamAF-RPN achieves precision rate of 0.808,ranking first among the compared algorithms,and it shows a bright antiinterference ability in the low resolution and aspect ratio change scenarios,proving the effectiveness of the proposed algorithm.