Modelling vehicle-pedestrian interactions at unsignalised locations employing game-theoretic models

There are some aspects of driver-pedestrian interactions at unsignalised locations that remain poorly understood. Understanding these aspects is vital for promoting road traffic safety in general which involves the interaction of human road users. Recent developments in vehicle automation have called for investigating human-robot interactions before the deployment of highly automated vehicles (HAVs) on roads so that they can communicate effectively with pedestrians making them trustworthy and reliable road users. To understand such interactions, one can simulate interactive scenarios studying various factors affecting road user decision-making processes through lab and naturalistic studies. To quantify such scenarios, mathematical models of human behaviour can be useful. One of these mathematical models that is capable of capturing interactions is game theory (GT). GT can provide valuable insights and strategies to help resolve road user interactions by analysing the behaviour of different participants in traffic situations and suggesting optimal decisions for each party. Thus, the current doctoral thesis aimed to investigate vehicle-pedestrian interactions at unsignalised crossings using GT models, applied to both lab-based and naturalistic data. One of the main aims of the current thesis was to understand how two or more human road users can communicate in a safe and controlled manner demonstrating behaviours of a game-theoretic nature. Thus, an experimental paradigm was created in the form of a distributed simulator study (DSS), by connecting a motion-based driving simulator to a CAVE-based pedestrian simulator to achieve this goal. It was found that the DSS could generate scenarios where participants interact actively showing similar communication patterns to those observed in real traffic. Another prominent finding was the stronger role of vehicle kinematics than personality traits for determining interaction outcomes at unmarked crossings, i.e. whether the pedestrian or driver passed first. To quantify the observations made from the DSS, five computational models namely four GT and one logit model were developed, tested and compared using this dataset. The GT models were obtained from both conventional and behavioural GT literature (CGT and BGT, respectively). This was done to bridge a gap in the previous research, specifically the lack of a comparison between these two modelling approaches in the context of vehicle-pedestrian interactions. Overall, the findings showed that: 1) DSS is a reliable source for the testing and development of GT models; 2) there is a high behaviour variability among road users highlighting the value of studying individualised data in such studies; 3) the BGT models showed promising results in predicting interaction outcomes and simulating the whole interaction process, when compared to the conventional models. These findings suggest that future studies should proceed to adopt, test, and develop BGT approaches for future HAV-human road user interaction studies. To validate the findings of the first two studies, a naturalistic study was conducted in the city of Leeds using state-of-the-art sensors. The sensors gathered road user data including their trajectory and speed over time. The findings from observations revealed similar communication patterns between drivers and pedestrians as in the DSS, suggesting a high degree of relative validity of the experimental paradigm. The results for the computational models were similar but the differences among the models were less noticeable compared to when the models tested against the controlled dataset. Overall, this thesis illustrates that the experimental paradigm and BGT models developed as part of the PhD programme have potential applications for HAV decision-making and motion planning algorithms, as well as traffic safety in general.