Research On Visual Object Tracking Method For Video Surveillance In Security

In the past,due to the insufficient computing capacity of the equipment,the security video surveillance system often used the strategy of sacrificing the performance of the algorithm to improve the running speed when deploying the visual tracking algorithm,which made it difficult for the existing technology to accurately track the target and give a security warning.Therefore,this thesis aims to study and design a visual target tracking algorithm that takes both performance and speed into account in the security video surveillance system to solve the above problems.Compared with the existing target tracking algorithms in security video surveillance,Siamese network based target tracking algorithm has higher accuracy and robustness,and can track various targets accurately according to different tasks,which can better meet the needs of security video surveillance.Therefore,in this thesis,we propose a siamese network based tracking algorithm based on adaptive search range adjustment and a lightweight siamese network based tracking algorithm in the visual target tracking framework.The main work of this thesis is summarized as follows:(1)The application of motion model based on spatiotemporal feature fusion to optimize the search range of tracking algorithm is studied.Aiming at the problem that the search center in the search range is liable to be punished by the cosine window and deviates from the real target,a search center optimization strategy combining the motion model based on recurrent neural network with the correlation filter response graph is proposed;Aiming at the problem that the search range can not include the target with large displacement,a search size optimization strategy based on inter frame velocity motion model is proposed.Experimental results show that these two search range optimization strategies can effectively improve the performance of siamese network based tracking algorithm in complex security scenes.(2)The influence of siamese network lightweight processing on mobile terminal deployment of tracking algorithm is studied.The applicability of the lightweight siamese network tracking algorithm in security deployment is verified by comparing the processing image speed,training difficulty,model size and tracking success rate of the algorithm backbone network before and after lightweight processing in the training process;By comparing the reasoning time and running memory occupation before and after the algorithm compression acceleration processing in the mobile terminal deployment process,the compression acceleration method that is most suitable for the lightweight siamese network tracking algorithm in the security deployment is verified.The experimental results show that although the tracking performance of the lightweight siamese network has declined to some extent,its speed and running memory usage have been greatly improved,which can meet the accuracy and real-time requirements of security deployment.