Research On Improved Siamese Network Single Target Tracking Method

Target tracking technology is one of the important research directions in the field of computer vision.In recent years,with the rapid development of deep learning technology and the increasing demand for applications,it has received widespread attention from researchers..Target tracking technology can be widely applied in fields such as intelligent monitoring,intelligent transportation,unmanned driving.It can achieve real-time tracking,recognition,and analysis of targets,providing strong support and assistance for various industries,and has broad application prospects.Although existing target tracking technologies have made significant progress,they still face the following challenges: 1)When facing complex scenarios such as large-scale deformation,complete occlusion,and background interference,the tracking target may drift or lose tracking;2)The backbone network used for feature extraction may lose some important features due to its deep layers,resulting in inaccurate tracking.In response to the above issues and difficulties,this article has conducted a series of research,and the main research work and achievements are summarized as follows.(1)A siamese network target tracking algorithm based on multiple branches(Siamese Network Target Tracking Algorithm Based on Multiple Branches,SiamMB)has been proposed.Firstly,a network robustness enhancement method is proposed by adding adjacent frame branches to improve the discriminative ability of target features in search frames,thereby enhance the robustness of the model.Secondly,by integrating spatial attention networks,different weights are applied to features at different spatial positions,with a focus on features that are beneficial for target tracking at spatial positions,thereby improving the model’s discriminative power.Finally,in the evaluation on the OTB2015 and VOT2018 datasets,SiamMB tracking accuracy and success rate reached 91.8% and 71.8%,respectively,achieving good competitiveness compared to current mainstream tracking algorithms.(2)A siamese network target tracking algorithm based on multi-layer feature fusion and ECA channel attention(Siamese Network Target Tracking Algorithm Based on Multilayer Feature Fusion and ECA-Net,SiamME)was proposed.Firstly,through a multi-layer feature fusion module,different levels of feature information are combined to enhance the robustness of the model and enhance its generalization ability.Secondly,a channel attention network is introduced to weight the information of each channel,thereby making the model pay more attention to important feature channels.Finally,evaluation was conducted on the OTB2015 and VOT2018 datasets,and experiments showed that the proposed SiamME tracking accuracy and success rate reached 91.5% and 70.6%,respectively,verifying the effectiveness of the model.(3)Finally,based on practical application requirements,a single target tracking prototype system was developed,and the target tracking model was encapsulated using the Django framework.A web system was developed using LayUI and SpringBoot technology.Implemented offline video file tracking and real-time online tracking functions.In addition,the model comparison function was achieved by simultaneously displaying the tracking effects of the two models.The system is deployed in the Docker application container engine environment and can run in different operating systems and hardware environments.