Modelling vehicle-pedestrian interaction behaviour and potential applications in connected and automated systems

Pedestrian safety remains a major concern in urban transportation systems, particularly as autonomous vehicles (AVs) are introduced into complex, mixed-traffic environments with diverse road-user behaviour and traffic characteristics. This thesis develops proactive safety frameworks to model vehicle–pedestrian interactions using trajectory data from multiple cities. Road-user behaviour is characterized using microscopic interaction variables, including speed, relative distance, and evasive actions. Data from different traffic environments, including Boston, Cairo, Las Vegas, Pittsburgh, Singapore, and Vancouver, are considered. The research addresses five related problems: (1) comparing the crash risk of AV–pedestrian and human-driven vehicle (HDV)–pedestrian interactions across environments; (2) modelling interaction behaviour using inverse reinforcement learning; (3) quantifying the degree of cooperation between road users in different traffic environments; (4) evaluating whether simulated conflicts reproduce the crash risk mechanisms observed in real data; and (5) applying these models to adaptive traffic signal control. Bayesian hierarchical extreme value models are used to estimate crash risk, while Multi-agent Adversarial Inverse Reinforcement Learning is used to model interaction behaviour. Results show that vehicle–pedestrian crash risk, evasive action behaviour, and cooperation vary substantially across environments, indicating that local traffic conditions play a central role in pedestrian safety. The comparison between the crash risk of AV–pedestrian and HDV–pedestrian interactions demonstrates key behavioural differences across various traffic environments. Also, transferring learned behavioural policies between cities altered interaction severity. Cooperation analysis showed varying levels of mutual adaptation during conflict avoidance strategies. Moreover, simulated interactions reproduced important behavioural patterns but frequently misrepresented crash risk magnitude, demonstrating the need for risk-based simulation calibration. Finally, the proposed safety-oriented adaptive traffic signal control framework using dynamic leading pedestrian intervals reduced crash risk and the number of conflicts, although vehicle delay increased to accommodate longer pedestrian times. Overall, this thesis contributes to the development of safer and more context-sensitive transportation systems by demonstrating that AV decision-making, safety-assessment tools, and traffic control strategies should account explicitly for local behavioural conditions, particularly in mixed and less-organized traffic environments.