FAM-LSTM: Predicting Macroscopic Pedestrian Dynamics Through Data-Driven Method

Crowd management in urban environments is increasingly crucial due to population growth and the resulting high-density pedestrian flows, which can lead to congestion and potential crowd disasters. Traditional continuum models for macroscopic pedestrian flow models usually assume that the pedestrian flow could be described by certain governing equations. Whether the dynamics of real crowd follow the governing equations remains unconfirmed, which would limit the ability of continuum models to reproduce real pedestrian dynamics. This paper introduces a novel data-driven approach, FAM-LSTM (Long Short-Term Memory with flow attention module) trained with empirical data from controlled pedestrian experiments, designed to predict macroscopic pedestrian dynamics. The flow attention module enhances the learning performance by accounting for the physical correlations between pedestrian density and velocity. Extensive testing of open-loop and closed loop predictions demonstrates that FAM-LSTM achieves satisfactory prediction accuracy, time adaptability, and robustness. The approach presented offers a beneficial advancement in macroscopic pedestrian dynamics prediction and provides a new perspective on crowd management efforts.

© 2025 IEEE. All rights reserved, including rights for text and data mining, and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission.