Design And Implementation Of Mobile Video Target Tracking Based On YOLO

Video target tracking technology is continuously booming,especially in the field of driverless and video surveillance.However,one of the problems that cannot be ignored is the long reasoning time and the increase of delay time caused by accessing remote servers in video target tracking,which greatly reduces the accuracy of tracking.In this paper,aiming at the above problems of target tracking,taking pedestrians as the research point,using DeepSORT tracking algorithm,YOLOv4 network as the detector,deployed on the embedded Jetson Nano platform,to build real-time target tracking system.In this paper,the storage structure of YOLOv4 target detection network and hardware platform had been optimized respectively.The main contents of this paper are as follows:Because the DeepSORT target tracking algorithm is based on the results of target detection,the target detector based on YOLOv4 algorithm was first trained.In order to obtain the deployment reasoning with low delay and high throughput and make the system achieve the goal of real-time detection,the ONNX and TensorRT algorithms were used to optimize the model by inter-layer or tensor fusion and multi-stream concurrent execution.By changing the power consumption mode and improving the heat dissipation of the system,the SSD solid-state hard disk with SATA3 interface was selected to optimize the storage structure and Swap partition of the platform to improve the system performance,so that the storage and cache capacity and speed of the system were improved,and the stability of the system were improved.Using TensorRT reasoning engine to optimize DeepSORT algorithm to achieve video target tracking.The experimental results showed that the video target tracking system built in this paper meets the practical application requirements of off-line,lower delay and more accurate.