Deep learning approach to force-based modeling of pedestrian flow in bottleneck scenarios

Abstract

This work introduces a hybrid force-based concept for modeling pedestrian flow through bottlenecks, combining deep learning with the Social Force Model. While the model dynamics is driven by the social-force based equation of motion, the driving forces are learned from trajectory data in the Bottleneck Caserne dataset using a neural network. Two concepts are analyzed: learning the force as one function (DirectForceNet) and learning the goal-directed, interaction, and environmental forces separately (FusionForceNet). Through data-driven simulations, both approaches have been shown to be applicable to bottleneck flow modeling. In addition to qualitative trajectory comparison, we conducted a quantitative evaluation against ground truth and the Social Force Model baseline. The DirectForceNet approach evinces the best quantitative comparison on testing dataset; however, the FusionForceNet demonstrates better ability to reproduce pedestrian dynamics in unseen synthetic scenarios.