Single Target Tracking And Application Based On Siamese Network

With the emergence of large-scale training datasets and the rapid improvement of computer’s data processing ability,the target tracking algorithm based on deep learning has achieved great success.As an important implementation method of deep learning target tracking algorithm,siamese network has been widely concerned because of its good tracking performance and good real-time performance.At the same time,due to the rapid development of UAV,the research of target tracking algorithm under the platform of UAV has great practical significance.But tracking algorithm based on siamese networks can be applied to uav platform tracking effect is not ideal,which is mainly due to the influence of UAV body structure,flight height,camera angle and other factors,resulting in the actual aerial video scene often has low resolution,scale change,deformation,occlusion,similar object interference,and so on and so forth.Therefore,how to build an efficient target tracking algorithm suitable for UAV platform is the focus of this paper.Aiming at the above problems,this paper will study the real-time target tracking algorithm based on siamese network.The main research work is as follows:(1)Research on target tracking algorithm of Siamese Network Based On Multi-layer Feature Adaptive Fusion(Siam MFAF).Aiming at the problem of poor tracking robustness in low resolution,scale change,occlusion,deformation and other scenes in aerial photography,a tracking algorithm based on Siamese Region Proposal Network(Siam RPN)is studied.Firstly,the feature extraction module of Siam RPN is improved,and a variety of modern deep coil models are replaced in Siam RPN.After comparing the tracking effects,an appropriate convolution model is selected as the feature extraction module.Secondly,the influence of deep and shallow convolution features on tracking is explored,and the Siam MFAF algorithm is proposed by combining the deep and shallow convolution with the dual attention mechanism model.The feature map of Siam MFAF algorithm not only has the high semantic information of deep convolution,but also has the spatial details of shallow convolution through the adaptive fusion of deep and shallow layer features by the dual attention mechanism model,so as to improve the robustness of tracking.Finally,in order to evaluate the tracking effect of Siam MFAF algorithm,an evaluation experiment is designed on UAV123 and DTB70 aerial aerial photography datasets.Experimental results show that compared with the benchmark Siam RPN,the tracking accuracy of the proposed algorithm Siam MFAF is improved by 3.9%and 6.3% respectively,and the success rate is improved by 4.2% and 9.3% respectively,which can effectively improve the tracking performance.(2)Research on target tracking algorithm of Siamese Cascaded Region Proposal Network Based On Multi-layer Feature Adaptive Fusion(CRPN-Siam MFAF).In order to solve the problem that the tracking accuracy of Siam MFAF algorithm decreases when the similar semantic objects interfere,the research of Siam MFAF tracking algorithm is carried out,and the CRPN-Siam MFAF algorithm is constructed.The algorithm performs more difficult negative sample sampling by cascading multi-order RPN,so that the number of positive and negative samples in each training sample of RPN network gradually tends to balance.Therefore,with the deepening of RPN stage,the classifier of RPN network has stronger recognition ability in order to distinguish complex interferences and targets.So as to improve the tracking accuracy and success rate.By designing evaluation experiments on UAV123 and DTB70 aerial aerial aerial photography datasets,CRPN-Siam MFAF tracking algorithm improved the tracking accuracy by 2.2% and 3.4%,and the success rate by 2.4%and 4.4%,respectively,compared with Siam MFAF algorithm.It can effectively solve the problem that Siam MFAF algorithm is easy to track drift when the similar semantic background is disturbed.