Research On Pedestrian Tracking And Pose Recognition Algorithm Based On Deep Learning

In the field of intelligent transportation,the goal of road has always been a hot research direction.Pedestrians are not only the main body of movement in road traffic,but also the main service object of vehicles.Therefore,the detection and recognition of pedestrians is the focus in traffic.With the development of computer vision technology,deep learning is more and more widely used in detection and recognition.In terms of identification equipment,there are also new technological innovations,from the original monocular camera hair to multi camera,as well as new lidar and point cloud to obtain the road information ahead.This paper studies pedestrian detection,pedestrian tracking and pedestrian pose recognition based on the images obtained by the on-board cameras of automatic driving and assisted driving:(1)In view of the complex background,large and dense number of pedestrians and the relative economy of the on-board hardware system,the expanded receptive field module is added to the Yolov3-tiny algorithm with fast detection speed to effectively improve the detection accuracy;The IOU value detection frame regression is changed to CIOU value detection frame regression with increased center point distance and aspect ratio,which can accelerate the model convergence and accurately locate the pedestrian position information;In the prediction stage of the algorithm,the detection results are filtered by transplanting the non maximum value and adding the IOU value of the center point distance,which avoids the phenomenon of violent elimination of the detection frame,and the algorithm is significantly improved in pedestrian occlusion and dense scenes.(2)In view of the instability and discontinuity of the detection frame in the detection algorithm,a pedestrian tracker is added to the pedestrian detector to make up for the defects of the detection algorithm.Based on the improved Yolov3-tiny pedestrian detector and the improved Deep SORT pedestrian tracker,this paper adopts TBD strategy to realize the real-time and accurate positioning of pedestrians.The lightweight shufflenetv2 is mainly used to retrain pedestrian appearance features for matching on Mars data set;DIOU is used to replace IOU for target matching.Finally,the tracker is adjusted according to the application scenario of this paper,and a pedestrian tracking algorithm with higher multi-target tracking accuracy is realized.(3)In view of the fact that the detection and tracking algorithm can not identify pedestrians with risk coefficient,it is proposed to add pedestrian attitude recognition and pedestrian distance measurement on the basis of pedestrian detection and tracking,which is conducive to improve the safety of automatic driving and auxiliary driving system,and produce feedback on pedestrian information in dangerous state and braking distance.In this paper,the Lightweight OpenPose algorithm is used to obtain the human body key map.Three methods are proposed to complete the pose recognition of pedestrians in four directions and three states according to the physical position of key points,the length width ratio of human detection frame connected by key points and the pose recognition network.Finally,based on the imaging principle of binocular camera,the pedestrian pose recognition and pedestrian ranging algorithm are proposed.Finally,the above methods are deployed on the embedded platform Jetson Nano to realize the implementation of the algorithm on the mobile terminal.The algorithm in this paper can actively promote the development of unmanned driving and intelligent driving technology,and the actual hardware application can also promote the development of industrialization.