Inferring and operating pedestrian behaviour models on autonomous vehicles

Interacting with pedestrians remains challenging for autonomous vehicles (AVs). In most current AVs, for safety and legal reasons, pedestrians are considered as obstacles, such that the AVs always stop for them. But their highly safe nature may lead pedestrians to take advantage over them and slow their progress.

When a pedestrian wishes to cross the road in front of the vehicle at an unmarked crossing, the pedestrian and AV must compete for the space, which may be considered as a game-theoretic interaction in which one agent must yield to the other. Game theoretic approaches have been used for decades to model the interactions between rational decision-makers, but have run in parallel streams with psychology research on human proxemics and trust. Results from game theory and psychology studies have yet to be operationalised for autonomous vehicles, this thesis thus aims to bridge the gap between these separate fields.

We first contribute with a comprehensive review of the literature in which we propose a new unifying taxonomy of pedestrian models required for autonomous driving, linking the low-level and high-level models of behaviour for the first time. We find that the low-level models are mature enough to be deployed in the real world but the high-level models such as game theory approaches still require more research and development. We therefore proceed with the evaluation of pedestrian interaction preferences with a game theoretic AV in a virtual reality experiment. Knowledge of such preferences could then be used by future AVs to predict and control for pedestrian behaviour. However, game theory approaches require the use of credible threats such as crash probabilities in order to make AVs progress on the roads, but another possible and more friendly solution that is explored in this work is Hall's theory of proxemics. Hence we propose a novel Bayesian method to infer pedestrian proxemic utility functions and the concept of physical trust requirement (PTR) for game theoretic AV interactions. We show how this PTR model can accurately generate Hall's empirical zone sizes, and then extend it to more general human-human and human-robot interactions. Finally, operating and deploying pedestrian behaviour models require the use of a physical AV platform, we hence introduce OpenPodcar, a low-cost, open source hardware and software platform developed for real-world AV research experiments with pedestrians. Thus, the present thesis forms a step towards the first operational game theoretic autonomous vehicles with pedestrian proxemic and trust behaviour.