Predicting the alighting behavior of metro passengers based on the analysis of action features

Abstract

Due to a lack of seat availability information regarding metro train carriages, passengers constantly move on platforms or inside carriages in search of seats, leading to overcrowding and safety concerns. Informing passengers of available seats and guiding them to less crowded carriages are beneficial for riding. The number of available seats at the next station is determined not only by the number of passengers currently occupying the carriage but also by the number of passengers alighting at the station. This study proposes a predictive framework that uses an optimized random forest model to estimate seat availability by analyzing metro passenger action features to predict alighting behavior. A dataset comprising 2009 passenger samples, encompassing 14 categories and totaling 9228 action features, was collected from video recordings in Changsha metro carriages. SHapley additive exPlanations (SHAP) analysis demonstrated that passengers’ actions as they approach the next station—particularly moving toward the carriage exit and standing up—are strong indicators for predicting alighting behavior. Certain combinations of actions, such as “watching the information display” and “standing up” showed superior predictive effectiveness compared with individual occurrences. The predictive model achieves high accuracy in assessing seat availability, contributing to improved service and boarding efficiency.