An Agent-Based Model for Trajectory Modelling in Shared Spaces: A Combination of Expert-Based and Deep Learning Approaches

Fatema T. Johora, Hao Cheng, Jörg P. Müller, Monika Sester

Abstract: Realistically modelling behaviour and interaction of mixed road users (pedestrians and vehicles) in shared spaces are challenging due to the heterogeneity of transport modes and the absence of classical traffic rules. Existing models have mostly used the expert-based approach, combining symbolic modelling and reasoning paradigm with the hand-crafted encoding of the decision logic. Recently, deep learning (DL) models have been largely used to predict trajectories based on e.g. video data. Studies comparing expert-based and DL-based micro-simulation of shared spaces concerning their accuracy are missing, and so are proven methodologies for combining these approaches into a single agent-based system. In this paper, we propose and compare an expert-based and a DL model and then combine them for trajectory prediction in shared spaces. Simulation results show the combined model to outperform both pure approaches in predicting realistic and collision-free trajectories.