Research On Multi-target Tracking Technology Based On Deep Learning

Multi-target tracking is a very important research direction in the field of computer vision.It has great academic research value and great commercial value,attracting more and more scholars,scientific research institutions and technology companies to research and develop multi-target tracking technology.Multi-target tracking is widely used in civil and defense fields.Multi-target tracking plays an important role in video surveillance and traffic monitoring in urban security and crowded places.It is an air defense early warning,weapon fire control system,missile guidance and emerging in recent years.One of the key technologies for UAV detection and firepower systems.Based on multi-target tracking of detection,the detection effect determines the tracking performance.Although the target detection based on deep learning greatly improves the detection accuracy,it still exists in the video frame or image sequence detection because of target motion,shooting angle,occlusion and other factors.The situation of missing the target.At present,the main research of multi-target tracking based on deep learning focuses on how to correlate the detected target information and prediction information,that is,the multi-target tracking is regarded as the correlation between the detection result and the prediction result,ignoring the problem that the target missed the target and the target is lost.Aiming at the above problems,this paper proposes a multi-target tracking method combining depth learning and feature point detection and matching under the multi-target tracking framework based on detection.The specific work is as follows:(1)Analyze the convolutional neural network and its components,comb the classic convolutional neural network model;analyze and experimentally compare two kinds of deep learning-based target detection algorithms,Faster R-CNN and YOLOv3;(2)The feature point detection and matching method is adopted to solve the problem of missed detection of the target detection module.The three feature point detection algorithms SIFT,SURF and ORB widely used in computer vision are analyzed and compared.The comprehensive feature point detection time,the number of features and the feature points are analyzed.Point distribution is used to select the appropriate feature point detection algorithm as the feature point detection algorithm of the multi-target tracking method proposed in this paper;(3)Completing the multi-target tracking method based on deep learning and feature point detection and matching,compared with the detection-based multi-target tracking method without adding feature point detection and matching,the tracking robustness of the proposed multi-target tracking method better.