Multiple Object Tracking for Robust Quantitative Analysis of Passenger Motion While Boarding and Alighting a Metropolitan Train

Abstract

To achieve significant improvements in public transport it is necessary to develop an autonomous system that locates and counts passengers in real time in scenarios with a high level of occlusion, providing tools to efficiently solve problems such as reduction and stabilization in travel times, greater fluency, better control of fleets and less congestion. A deep learning method based in transfer learning is used to accomplish this: You Only Look Once (YOLO) version 3 and Faster RCNN Inception version 2 architectures are fine tuned using PAMELA-UANDES dataset, which contains annotated images of the boarding and alighting of passengers on a subway platform from a superior perspective. The locations given by the detector are passed through a multiple object tracking system implemented based on a Markov decision process that associates subjects in consecutive frames and assigns identities considering overlaps between past detections and predicted positions using a Kalman filter.