Metro Passenger Flow Statistics Based on YOLOv3

Abstract

Passenger flow real-time detection in per train compartment is the basis for realizing metro entrance flow restriction, platform waiting guidance and optimizing train departure interval. In this paper, we collect RGB images as the input of the improved YOLOv3 network, then we can get the statistics of passenger flow by detecting and tracking passengers’ heads. For the head detection part, this paper re-selects the number of anchors by using the k-means++ clustering method and removes the large-scale object detection module in the original YOLOv3 network. For the head tracking part, this paper proposes the probability gradient map, IOU-Max maximum matching algorithm and voting strategy, which make the stability of target tracking greatly enhanced. Compared with the original YOLOv3 network, our research results show that the accuracy of head detecting based on the improved YOLOv3 network increases from 90.1% to 97.84%, the recall rate increases from 85.4% to 94.7%, and the accuracy of the head tracking algorithm is as high as 98.87%, and the accuracy of passenger flow density detecting is as high as 95%. The measured detection speed reaches 47FPS on the TITAN X server, enabling real-time fast detection and tracking of targets.