Research Of Pedestrian Tracking System Based On YOLO V5 And Deep SORT

Target tracking technology is one of the research directions in the field of computer vision,which plays an important role in smart city,intelligent transportation and other intelligent industries.With the rapid development of deep learning,great breakthroughs have been made in the difficult problems that traditional algorithms are difficult to solve,which promotes the rapid development of machine vision.Most of the existing multi-target tracking algorithms are based on detection,and the detection result of target detector directly affects the tracking performance of the algorithm.In actual monitoring scenarios,pedestrian detection often encounters problems such as illumination change,occlusion,scale change,target movement and so on,and is prone to miss and misdetection,especially when the target is small,resulting in poor detection performance.When the tracking target has occlusion,tracking errors and ID-switch problems often occur in the process of target tracking.Therefore,it is urgent to solve the problem of how to build a robust multi-target tracking system.In this thesis,target tracking is divided into two sub-tasks: target detection and target tracking.In target detection,an efficient and accurate detector is found to detect pedestrians in surveillance videos.In target tracking task,target tracking is completed through data association of detection results of the detector.The main work of this thesis is as follows:(1)On the basis of existing studies,MOT16,the popular pedestrian multi-target tracking data set,was selected.By comparing deep residual network,Faster-RCNN detector and YOLO detector series,YOLO v5 detector with excellent performance was finally selected as the front detector of pedestrian tracking system.At the same time,a recognition system method based on feature extraction technology of Gaussian cosine curve in wavelet transform and Distractor classifier is proposed to improve the detection result of YOLO v5 model.Finally,the detection precision is high and the accuracy is improved on public and homemade data sets.(2)Deep SORT is used to complete the tracking of pedestrians.Aiming at the problem of pedestrian ID hopping caused by occlusion and other reasons in the tracking system,an ultra-short distance positioning algorithm based on multi-dimensional intersection ratio and extended cascade matching method is proposed to improve the matching effect of the algorithm.The improved Deep SORT tracking algorithm can effectively reduce the false detection rate and improve the recognition speed and accuracy.Experiments show that the scheme can effectively improve the positioning accuracy and track the target in a large range,which verifies its practicability.(3)Combined with the improved YOLO v5 algorithm and Deep SORT algorithm,YOLO v5 is used as the detector of Deep SORT target tracking,and a pedestrian detection and tracking system based on daily surveillance video is designed.Detailed requirement analysis and design are made for the system.After completing the system design,Functional and non-functional tests were carried out on the system.The accuracy and stability of the system met the expected requirements,and it could complete the daily pedestrian detection and tracking work.