Subway Platform Passenger Flow Detection Based On Machine Vision

With the acceleration of urbanization,the subway has become the core component of urban rail transit.Due to the large passenger flow of the subway,it poses a serious challenge to ensure the safe operation of the subway.At present,subway passenger flow detection is mainly based on gate data,but the distribution and flow direction of passenger flow in the station cannot be obtained in this way.This paper studies the passenger flow detection method of subway station platform based on machine vision,which is of great significance for subway emergency management and intelligent scheduling.By analyzing the network structure of the existing models,the head detection model based on detection and regression methods is realized,and the two models are improved and compared.Aiming at the problem that the single camera coverage is small and the dual camera has the problem of overlapping field of view,a passenger flow detection algorithm based on the dual camera is designed;A passenger flow detection algorithm suitable for subway scenarios is designed,which solves the problem that the traditional passenger flow detection algorithm cannot detect multiple doors at the same time and the accuracy rate is low.Finally,the spatio-temporal distribution characteristics of passenger flow of subway stations are studied.The specific contents are as follows:1.Based on the detection method and the regression method,the head detection model is realized.The detection method is based on the YOLOv5 x model,and the m AP reaches 95.8%on the self-built human head dataset.A new YOLOv5x-SDH model is built by adding the small object detection module in YOLOv5 x model,and the m AP increased by 1.1%.The regression method is based on the Focal Inverse Distance Transform map(FIDTM),and a new FIDTM-LFA model is built by using the Layer-by-Layer Feature Fusion Module(LFA).Experimental results show that the FIDTM-LFA model has higher detection accuracy in population counting and localization tasks.Finally,the detection performance and applicable scenarios of the two models are analyzed.2.A dual-camera-based passenger flow detection algorithm(DPFD)is presented.The DPFD algorithm obtains the overlapping domain between the dual cameras through four steps of SIFT feature point extraction,KNN feature point matching,RANSAC algorithm filtering and homography matrix,and the overlapping domains are remediated in the pair installation scenario.Based on the YOLOv5x-SDH model and ray method,the passenger flow in the dual camera scenario is counted.Experimental results show that the DPFD algorithm has achieved better detection results than the dual-camera direct counting method.In the two types of dual camera parallel and opposite installation scenarios,the accuracy rate of 90.7% and 94.6% was achieved,respectively.3.A passenger flow detection algorithm(MDDC)based on the direction of movement of the target and collision detection is given.Based on the YOLOv5x-SDH model and the Deepsort algorithm,the MDDC algorithm judges the loading and exit behavior by whether the target has reached the door position(collision detection area)and the direction of the target’s motion.Compared with the baseline method,MDDC can detect multiple doors at the same time,and the detection accuracy rate reaches 89.1%.The DPFD algorithm can obtain the safety level of each area of the subway platform according to the detection results,and the MDDC algorithm can obtain the passenger flow of the entire train.The subway station managers flexibly formulate emergency management plans according to the characteristics of passenger flow time and space distribution,which can effectively improve work efficiency and emergency management level.