Artificial Neural Networks Predicting Pedestrian Dynamics in Complex Buildings

Abstract

The prediction of pedestrian movements in complex buildings is a difficult task. Recent experiments have shown that the behaviour of pedestrians tends to depend on the type of facility. For instance, flows at bottlenecks often exceed the maximal rates observed in straight corridors. This makes pedestrian behaviours geometry-dependent. Yet the types of geometries are various, and their systematic identification in complex buildings is not straightforward. Artificial neural networks are able to identify various types of patterns without supervision. They could be a suitable alternative for forecasts of pedestrian dynamics in complex architectures. In this paper, we test this assertion. We develop, train and test artificial neural networks for the prediction of pedestrian speeds in corridor and bottleneck experiments. The estimations are compared to those of an elementary speed-based model. The results show that neural networks distinguish the flow characteristics for the two different types of facilities and significantly improve the prediction of pedestrian speeds.