Abstract

Autonomous Vehicles (AVs) must interact with other road users. They must understand and adapt to complex pedestrian behaviour, especially during crossings where priority is not clearly defined. This includes feedback effects
such as modelling a pedestrian’s likely behaviours resulting from changes in the AVs behaviour. For example, whether a pedestrian will yield if the AV accelerates, and vice versa. To enable such automated interactions, it is necessary for the AV to possess a statistical model of the pedestrian’s responses to its own actions. A previous work demonstrated a proof-of- concept method to fit parameters to a simplified model based on data from a highly artificial discrete laboratory task with human subjects. The method was based on LIDAR-based person tracking, game theory, and Gaussian process analysis. The present study extends this method to enable analysis of more realistic continuous human experimental data. It shows for the first time how game-theoretic predictive parameters can be fit into pedestrians natural and continuous motion during road-crossings, and how predictions can be made about their interactions with AV controllers in similar real-world settings.