[Paper Review] Towards predicting Pedestrian Evacuation Time and Density from Floorplans using a Vision Transformer

Patrick Berggold, Stavros Nousias+2 more|arXiv (Cornell University)|Jun 27, 2023

Evacuation and Crowd Dynamics4 citations

TL;DR

This paper proposes a Vision Transformer-based deep learning model that predicts pedestrian evacuation time and spatiotemporal density heatmaps directly from floorplans, bypassing traditional simulation bottlenecks. Using a synthetic dataset generated via a parametric pipeline with a conventional simulator, the model achieves fast, accurate predictions and is integrated into a BIM tool for real-time design feedback.

ABSTRACT

Conventional pedestrian simulators are inevitable tools in the design process of a building, as they enable project engineers to prevent overcrowding situations and plan escape routes for evacuation. However, simulation runtime and the multiple cumbersome steps in generating simulation results are potential bottlenecks during the building design process. Data-driven approaches have demonstrated their capability to outperform conventional methods in speed while delivering similar or even better results across many disciplines. In this work, we present a deep learning-based approach based on a Vision Transformer to predict density heatmaps over time and total evacuation time from a given floorplan. Specifically, due to limited availability of public datasets, we implement a parametric data generation pipeline including a conventional simulator. This enables us to build a large synthetic dataset that we use to train our architecture. Furthermore, we seamlessly integrate our model into a BIM-authoring tool to generate simulation results instantly and automatically.

Motivation & Objective

• To address the computational inefficiency and complexity of conventional pedestrian simulators in building design.
• To develop a data-driven approach that predicts evacuation time and density heatmaps directly from floorplans.
• To overcome the limited availability of real-world evacuation datasets by generating a large-scale synthetic dataset using a parametric pipeline.
• To integrate the model into a BIM-authoring tool for real-time simulation feedback during design.
• To demonstrate that deep learning can match or exceed simulation accuracy with significantly reduced runtime.

Proposed method

• A Vision Transformer architecture is trained to predict spatiotemporal density heatmaps and total evacuation time from floorplan images.
• A parametric data generation pipeline simulates pedestrian evacuations using a conventional simulator to create a large synthetic dataset.
• The model is trained end-to-end on floorplan images with corresponding evacuation time and density map annotations.
• The architecture processes floorplans as input tokens, leveraging self-attention mechanisms to capture spatial dependencies across the layout.
• The model is deployed within a BIM environment to enable instant prediction generation during design workflows.
• Loss functions are optimized for both evacuation time regression and pixel-wise density heatmap prediction.

Experimental results

Research questions

• RQ1 Can a Vision Transformer model accurately predict pedestrian evacuation time from floorplans without running full simulations?
• RQ2 To what extent can a deep learning model generalize to predict spatiotemporal pedestrian density distributions from floorplan layouts?
• RQ3 How effective is a synthetic dataset generated via a conventional simulator in training a generalizable evacuation prediction model?
• RQ4 Can the model be seamlessly integrated into BIM tools to enable real-time design feedback?
• RQ5 Does the proposed model achieve comparable or better performance than traditional simulators with significantly reduced inference time?

Key findings

• The model achieves high correlation between predicted and ground-truth evacuation times across diverse floorplan configurations.
• The predicted density heatmaps accurately reflect congestion patterns and evacuation dynamics observed in simulation.
• The model reduces inference time from minutes (in traditional simulators) to milliseconds, enabling real-time feedback in design workflows.
• The synthetic dataset generation pipeline effectively captures realistic evacuation behaviors, enabling generalization to unseen layouts.
• Integration into a BIM tool enables instant simulation results, significantly accelerating the design iteration process.
• The Vision Transformer outperforms baseline models in both evacuation time prediction and density heatmap generation accuracy.