Research On Recognition Method Of Pedestrian Bottleneck Identification In Subway Station Based On Deep Learning

Subway stations are usually characterized with large volumes of passengers,relatively closed space,and natural bottlenecks such as turnstiles,all of which necessitates the needs for monitoring the pedestrians in the stations.With the advanced development of deep learning technology,a number of methods have been proposed to detect and track pedestrians efficiently and accurately,providing new insights for identifying bottlenecks.The bottleneck studied in this paper refers to the state of high density and low speed of pedestrians.This article is dedicated to applying deep learning-based target detection and tracking algorithms to achieve pedestrian detection and tracking,and further calculating pedestrian density,flow rate,speed.Based on these indicators,pedestrian bottleneck can be specified.In this paper,YOLOv3 and Deep SORT algorithms are used to realize pedestrian detection and tracking at the turnstiles in subway stations.Based on the surveillance video at the turnstile,a pedestrian sample data set was constructed,and the YOLOv3 model was trained on the constructed data set by transfer learning using 2080 Ti GPU.Then the model weight invoking the smallest loss function value of the verification set was saved.The trained detection model combined with the Deep SORT algorithm is applied to realize pedestrian tracking.Relying on the results of pedestrian detection and tracking,the calculation method of pedestrian density,flow rate and speed at the gate is proposed.The speed calculation is performed using a vanishing point-based camera calibration at the turnstile.The surveillance videos at the turnstile for part of the time period was selected to verify the applicability of the proposed methods: Based on the precision,recall and AP obtained from the YOLOv3 algorithm,it is verified that the algorithm has a strong performance in pedestrian detection.Calculate and analyze the pedestrian flow characteristics in terms of density,flow rate,and speed of the pedestrian.Furthermore,fuzzy set theory is used to determine the level of pedestrian bottleneck.Taking density and flow rate as the evaluation index of the bottleneck level at the turnstile.Take one-hour video as an example to calculate the density and flow rate of pedestrians in different time periods.According to pedestrian density and flow rate,the bottleneck level is determined to be "non-bottleneck".Furthermore,the passage characteristics of pedestrians are analyzed,including trajectory and utilization of the turnstile.It proves that there are differences in the usage of the four passages of the turnstile.Most pedestrians will choose the passage with the shortest walking distance to enter the station,and one of the passages is used the least frequently.Besides,the rationality of four passages in the turnstile are evaluated in terms of facility passing capacity and pedestrian passage characteristics,respectively.It is concluded that the passage with the lowest frequency could be cancelled and the other three should be remained.A pedestrian bottleneck identification system for subway stations is built using Flask as a framework,which includes three modules: view,controller,and model,with functions such as video upload,pedestrian detection and tracking,pedestrian density and flow rate calculation,and result display.Part of the video is selected for presentation,and it is demonstrated that all functions of the system can be used normally,so that it can be straightforward to realize the monitoring of pedestrian traffic as well as the identification of traffic bottlenecks.