The study of self-organised behaviours and movement pattern of pedestrians during fire evacuations: virtual experiments and survey

Abstract

In this study, Minecraft was adopted to investigate pedestrian behaviours and evacuation processes in non-fire (normal) and fire conditions involving multiple simultaneous players. We built a PaddlePaddle-based deep learning platform for the automatic recognition of pedestrian coordinates from the Minecraft background database. To verify the virtual experiments in Minecraft, we conducted a series of experiments under identical conditions in a real (physical) room and virtual environment on the Minecraft platform. The pedestrian behaviour, flow patterns and evacuation time in the Minecraft experiments fit well with the results of the real-life experiments, indicating the potential of Minecraft for realistically simulating evacuation processes. To further demonstrate the feasibility of using Minecraft for evacuation studies, we conducted a series of virtual evacuation experiments with different fire configurations and room geometries in Minecraft. The pedestrians were found to keep higher distance towards fire and greater detour angle during the fire emergency in a square room. Furthermore, in the scenario of a fire in a corridor, in addition to the aforementioned fire avoidance behaviours, the pedestrians exhibited orderly queuing and exited the room one by one, resulting in increased walking speeds and thus a faster evacuation for most pedestrians. Finally, one post-experiment survey was designed to assess participants' perceptions of the Minecraft environment and the pedestrian behaviours and related psychological factors. The results show that the Minecraft can generate considerate realistic real-life situations, and pedestrians chose self-organised behaviours including away from fire, along the wall and detour towards fire in both survey and evacuation experiments.