Research On Pedestrian Detection And Tracking Technology In Traffic Hub Based On Improved YOLO V3

As one of the basic components of auxiliary driving system,vehicle monitoring system and early warning protection system,pedestrian detection plays an important role in many fields,especially in the various crowded environments of transportation hubs.Moreover,pedestrian detection technology plays an important role in traffic safety,public service and data processing as the basic support of passenger flow statistics,passenger flow guidance and safety early warning.Aiming at the problems of traffic hubs’ pedestrian density,serious occlusion and low detection precision,this thesis proposes a detection algorithm based on the improved YOLO v3,with the idea of head--shoulder detection.The main contents are as follows:(1)In order to deal with the lack of pedestrian data set in transportation hubs,this thesis produces a head--shoulder data set for pedestrians at the transportation hubs,including the scenes of train exit,traffic intersection,station square,etc.A total of 6196 pictures are recorded,and the number of marked valid head--shoulders is 83072,and the data set is named as Transportation＿hub＿Human.In order to verify the robustness of the algorithm,the public data set Crowd Human is selected for testing.In this thesis both Transportation＿hub＿Human and Crowd Human are processed into standard VOC data set format,for the convenience of training and testing.(2)As for the YOLO v3 algorithm,three improvements are proposed.Firstly,in view of the problem that the K-means clustering algorithm depends heavily on the initial value,the K-means++ algorithm is used to optimize the K-means algorithm to gain better prior bounding box parameters and improve the coincidence degree between the prior bounding box and the data set.Secondly,the width of the network is widened and the depth of the network is reduced to enlarge the visual field threshold for the detection of small targets to avoid the gradient disappearing.Finally,a detection layer is added on the original three detection layers to increase the detection accuracy of small targets.Through training and validation on Transportation hub＿Human and Crowd Human,the experiment shows that the improved algorithm increases the detection accuracy significantly.(3)The realization of pedestrian tracking is based on DeepSORT algorithm.The improved YOLO v3 algorithm is trained on the produced data set,and the model obtained is used as the detector of the DeepSORT algorithm.Experiments show that the improved algorithm has a higher tracking accuracy.A pedestrian tracking software based on improved YOLO v3 algorithm is designed and developed,with the upper machine making it more convenient for users to operate.The software has the function of pictures and videos detection,and can display the number of people detected in real time.This thesis studies the algorithm of pedestrian detection and tracking at traffic hubs based on improved YOLO v3,takes the model trained by the improved YOLO v3 algorithm as the detector of the DeepSORT algorithm,overcomes the problem of low accuracy in detection and tracking of traditional model,and realizes the effective monitoring of pedestrians at traffic hub.The related research has good reference value for academic research on intelligent auxiliary driving system,vehicle monitoring system and early warning protection system.