Spatio-temporal Multi-scale Pedestrian Flow Model by using Attention Module

This is the authors' abstract. We don't add key points for this paper.

Congestion in urban spaces sometimes leads to serious accidents. To prevent a congestion and serious accident, it is important to properly arrange building-layouts and guide pedestrians safely based on mathematical modeling and simulation of pedestrian flows. In the real world, however, each pedestrian’s behavior is so complicated and often includes spatio-temporally multiscale decision making while most of existing studies mainly focused on a limited local area. In a previous report, a trained model could simulate some pedestrian trajectory, but it also generated undesirable behavior that a pedestrian ignored a traffic light. In this study, we introduce an attention mechanism to solve this problem. The performance of the model is evaluated quantitatively by the prediction error and qualitatively by visualizing the time of a traffic light switching.