Software Devclopment Of Person Re-identification And Tracking System Based On Tegra Xavier

Video surveillance system can monitor and protect an area in real time.However,with the increasing demand of security under various scenarios and with the increasing of the number of surveillance cameras,real-time monitoring can barely meet the requirements of modern intelligent video surveillance systems.Prisons are important facilities of the national judiciary,the video surveillance system of which requires high security.It needs to identify and track persons across multiple cameras.Therefore,it is of great significance to study and develop an intelligent video surveillance system which is able to achieve the real-time and accurate detection,re-identification and tracking of persons.In view of the shortcoming that the traditional video surveillance system of prisons lacks the capability of intelligent analysis,this thesis studies and develops a person re-identification and tracking system software based on embedded GPU,Tegra Xavier.The whole system is divided into three modules.The streaming media processing module receives and decodes the input network video stream,as well as encodes the result images into video stream and transmits it.The system function module realizes the communication between the processing terminal and PC monitoring platform,distributes multiple input video streams,and records the system working status.The algorithm processing module uses the improved SSD algorithm to detect persons,uses the person re-identification algorithm to extract features of persons and track them across multiple cameras,and uses the DeepSORT algorithm to achieve multi-person tracking.The system developed in this thesis takes full advantage of embedded GPU and TensorRT to optimize algorithms based on convolutional neural network.According to the experimental results,the system of person re-identification and tracking based on Tegra Xavier realizes real-time and accurate person detection,re-identification and multi-camera tracking.The processing time of person detection is 9.00 milliseconds for one frame,while the processing time of feature extraction is 16.85 milliseconds for one frame.The average tracking accuracy is above 93%.In summary,the system shows high value of engineering application.