Research On Improved YOLOv3 Algorithm For Pedestrian Detection And Tracking

Pedestrian detection and tracking can provide important information for the investigation of abnormal personnel,and can make some general statistics on the traffic volume of the street.This is quite practical in the current epidemic environment,and it can also detect the densely populated areas,and provide useful data for traffic safety and social public safety management.With the development of science and technology,the computer hardware upgrading makes the pedestrian detection and tracking method based on deep learning widely used,but there is still some room for improving the accuracy of pedestrian detection and tracking.How to detect and track the pedestrian on the road is the focus of this paper.In the real world,some of the surveillance cameras are very close and some are very far away,the obstructions may appear on the road,and pedestrian as a specific detection target has its unique appearance characteristics.Therefore,in this paper,for the pedestrians in road surveillance video,the improved loss function and residual module of YOLOv3 network combined with Deep Sort is proposed to detect and track pedestrians,and then make relevant comparative experiments and analyze and summarize the experimental results.The research work of this paper is divided into two aspects:(1)A pedestrian detection network based on YOLOv3 is proposed,which improves the loss function and residual module.First,improve the loss function of YOLOv3 based on the size and proportion of pedestrians in the real environment,and propose the DPIOU(Distance and Proportional-IOU)loss based on the IOU loss,the DPIOU loss added penalty items based on pedestrian characteristics,solved the problem that IOU loss does not produce sliding gradients when the detection frames do not intersect,speeds up the iterative speed of the network,and increases the accuracy of the detection frame coordinates;secondly,the Focal Loss is proposed to solve the problem of the imbalance of positive and negative samples in the YOLOv3 network;then the residual module of the feature extraction network is optimized to reduce the loss of feature information and improve the overall performance of the network.Finally,through experiments,it is concluded that the convergence speed of this network is faster than the network before the improvement,and the accuracy is improved by 10%.(2)Use the improved YOLOv3 as a pedestrian detection network before tracking,combined with Deep Sort to track pedestrians,and make traffic statistics for pedestrians in the current environment.First,the improved YOLOv3 network is used to improve the accuracy of Deep Sort input data,and the improved YOLOv3 as a one-stage detection network meets the real-time detection;secondly,aiming at the problems of occlusion on the road and the occlusion of pedestrians,a method based on the appearance feature matching of pedestrians is proposed to improve the network's ability to track occluded pedestrians;then,use the trajectory intersection count to count the pedestrian flow.Finally,through experiments,it can be seen that the accuracy of the improved tracking network is significantly improved,the loss of blocked pedestrians is reduced,and the accuracy of traffic statistics is also improved by 3.5%.