Predicting occupant evacuation times to improve building design

Building design requires considering multiple requirements and must fulfill diverse regulations. Therefore, model analysis and simulations are fundamental parts of the design process to find the optimal solution for a given problem. Important decisions are based on a building’s assessed final performance in the early design phases. In particular, the analysis of pedestrian flow dynamics is paramount for public facilities like train stations concerning occupants’ comfort and evacuation behavior. Currently, it requires multiple steps, from preparing the BIM model to performing pedestrian flow analysis, including semi-automated, often manual work that demands high computation times. Therefore, to improve the building design efficiency in terms of time and pedestrian circulation, this paper proposes a framework applying Deep Learning methods. We propose a real-time pedestrian evacuation prediction to replace time-consuming pedestrian dynamics simulations. More precisely, a modular neural network architecture is designed, including a Convolutional Neural Network and a Multilayer Perceptron, that takes floorplan images and building and simulation parameters as input and predicts the crowd evacuation time for a given building model. As a result, a mean prediction accuracy of 15% could be achieved.