An extended cellular automaton model for crowd evacuation under multi-storey building with ControlNet

Abstract

Vertical evacuation safety in high-rise buildings presents a key challenge for urban resilience. This study proposes an automated evacuation modelling method for high-rise buildings that combines deep learning and an extended cellular automaton model, which can achieve rapid and reasonable evacuation modelling under customized multi-layer building scenarios. A ControlNet is integrated to convert building floor plans into semantic feature maps, and a multi-level cellular automaton framework is constructed that includes floor layouts and bilateral stairwells, allowing to customize the number of floors and visualize dynamic evacuation process between staircases. Comparative analysis with validated models and actual evacuation drill data, the proposed method shows a higher semantic segmentation accuracy (IoU=0.906) and more accurate evacuation time prediction (Error<9%). Moreover, the proposed method automates the semantic interpretation of floor plans, enabling the "image-to-simulation" automation and the generation of high-rise simulation scenarios directly from images within minutes, while effectively capturing the merging effect. The analysis also indicates that the number of stairwells and their internal width have a decisive influence on overall evacuation efficiency. This study aims to provide an efficient tool for the intelligent transformation of performance-based evacuation design and emergency management in high-rise buildings.