Cognitive Mechanisms of Behaviour Estimation: Modelling Pedestrian Interpretation of Approaching Vehicle Behaviour

Understanding and modelling the interactions between pedestrians and automated vehicles (AVs) is important for facilitating widespread and safe AV deployment. During pedestrian-vehicle interactions, pedestrians form and update their beliefs regarding the vehicle’s behaviour as it approaches. The mechanisms determining how a pedestrian interprets the behaviour of an approaching vehicle remain unclear. Previous studies have proposed models of cognitive mechanisms, such as estimating the goals of other agents, but none has attempted to apply and model the behaviour estimation mechanism in the real and dynamic context of pedestrian-vehicle interaction. Drawing inspiration from cognitive science, a first vehicle behaviour estimation experiment was conducted, and existing Bayesian observer models of goal estimation were modified and applied to the pedestrian crossing setting. Thus, an observation-based model with two alternatives based on 1) direct deceleration perception and 2) a more plausible visual cue, the rate of change of the relative rate of optical expansion tau-dot, was proposed. The first experiment demonstrated that pedestrians do not solely rely on deceleration-related cues to judge whether an approaching vehicle is stopping, but that the vehicle’s kinematic conditions, specifically its speed, time-to-arrival, and overall manoeuvre time history, also influence their beliefs. Even though the observation-based model achieved a relatively high correlation between model predictions and average pedestrian beliefs, it did not predict all the average pedestrian belief patterns in detail, being quite limited in predicting beliefs when the vehicle maintains constant speed. So, it was assumed that pedestrians may be utilising prior knowledge and situational expectations when the vehicle is far away, while deceleration observations become more crucial as the vehicle approaches. Thus, pedestrians likely infer the vehicle’s behaviour by both directly observing the vehicle’s actions and expecting the driver/AV to follow the most beneficial (value-maximising) behaviour. This rational, value-maximising reasoning mechanism was proposed as the value-based model. Both observation-based (Ob) and value-based (Vb) models were then integrated into an augmented model (Ob+Vb). All three models were evaluated for their ability to predict average pedestrian beliefs regarding the approaching vehicle’s behaviour. This evaluation illustrated that Ob struggled with constant speed scenarios due to its reliance on deceleration cues, Vb captured most kinematic effects but had limitations when the approaching vehicle is close to stopping, and Ob+Vb leveraged the strengths of both previous models, accurately reflecting all kinematic effects and belief patterns, achieving near-perfect correlation and the lowest error. Finally, to validate the three behaviour estimation models and test their generalisability, a second vehicle behaviour estimation experiment was designed and conducted, and the models’ predictive capabilities were evaluated on the resulting dataset. Analyses on this dataset demonstrated the replication of previous findings in identical kinematic scenarios, validating the models, particularly Ob+Vb, which accurately predicted pedestrian beliefs, again. Furthermore, Ob+Vb successfully generalised to unseen scenarios with varied speeds and new manoeuvres, showing its ability to predict beliefs in novel situations. Additionally, Ob+Vb again exhibited superior performance, obtaining near-perfect correlation and the lowest error compared to the other models. Together, these studies demonstrate that: 1) while Bayesian observation of behaviour may suffice for simple laboratory tasks, it falls short in real traffic contexts, 2) pedestrians assess approaching vehicle behaviour by combining observations of vehicle actions with expectations of the driver's most rational, value-maximising future actions and 3) the proposed augmented model successfully predicts pedestrian beliefs, reproducing findings quantitatively and qualitatively, illustrating generalisability, and providing a likely explanation of the mechanisms with which a pedestrian interprets the behaviour of an approaching vehicle. Overall, investigating and modelling behaviour estimation in the pedestrian-vehicle interaction setting and its underlying mechanisms, present a significant challenge. However, this thesis demonstrates that it is possible to gain deeper insights into how a pedestrian interprets an approaching vehicle’s behaviour, by integrating different psychological theories. This thesis not only enhances the theoretical understanding but also offers practical implications for designing safer and more intuitive interactions between pedestrians and AVs.