Design Of Moving Object Recognition And Tracking System Based On Deep Learning

The development of deep learning technology has made great progress in the field of artificial intelligence,especially in computer vision and vision.Target tracking is an important research field in computer vision,and it has important application value in intelligent monitoring,automatic driving and other dual-use areas.Although some researches have reached the top precision,the research results prove that deep learning has the same advantages in the field of target tracking,but it is difficult to achieve the real-time requirement,and the traditional method is still used in the project.In this paper,deep learning target recognition tracking model is applied to Jetson tx2.Firstly,the real-time tracking accuracy of object is improved by combining YOLOv3 model with HOG feature and Deepsort method,then,the YOLOv3 network is simplified and integrated with Deepsort and ported into Jetson tx2 to realize the real-time tracking function of the target recognition and tracking network in the embedded system.First of all,this paper introduces the basic concepts of deep learning and the related techniques of target learning and target tracking in deep learning,this paper introduces two kinds of models based on regression and candidate area in deep learning field,and selects the classical network models to explain,and compares the structure of SSD and Yolo series with the actual data set test Then,in order to improve the accuracy of the deep learning network in the project of target recognition and tracking,we propose a method to fuse the HOG feature in the deep learning feature,considering the computing resource limitation of embedded platform,the tradeoff between accuracy and speed is made,and a deep learning based object recognition and tracking system is built on the low power embedded platform of NVIDIA tx2,which has GPU,a target tracking model suitable for embedded engineering project is trained as the core method of the system,and the performance of the system is tested by experiment.The experimental results show that the newly designed embedded system can realize the real-time recognition and tracking of the target,and the position information of the object is similar to that of the original network.