Real-time monitoring model of passenger flow congestion of subway station based on improved YOLOv3

Abstract

The monitoring of passenger flow congestion in subway stations is the key to passenger flow analysis and safety control of rail transit. For the problem of passenger flow monitoring in subway station, a real-time monitoring method of passenger flow congestion based on surveillance video is proposed in this paper. Considering the characteristics of high density and serious occlusion of passenger flow in subway stations, this paper constructs a real-time monitoring model of passenger congestion in subway stations based on improved YOLOv3 is built. Firstly, YOLOv3 algorithm is modified as follows: CIoU is introduced instead of IoU to calculate the regression loss function of the target bounding box, which improves the positioning accuracy of the target bounding box; the Focal Loss function is added to guide the model to focus on the training of hard-to-class samples with high occlusion; the K-means clustering algorithm is introduced to re-cluster on the self-made dataset to obtain the anchor boxes suitable for subway station passenger flow. Then the congestion level of different areas in subway stations is divided, and the method of passenger flow congestion identification is given. Finally, an experiment is carried out based on the self-made subway pedestrian dataset SOP. The experimental results show that the average accuracy of this method increases from 92.11 % to 94.45 %, the average detection speed increases from 28.634 fps to 29.681 fps compared to former algorithm, which can meet the requirements of real-time and accuracy of pedestrian detection in subway stations.