Evaluating unsignalized crosswalk safety in the age of autonomous vehicles

As autonomous vehicles are poised to enter public roadways, a major concern is their interaction with pedestrians. It requires attention and ability for pedestrians to interact correctly and for autonomous vehicles to detect pedestrians hence avoiding collisions. We propose a complete pipeline to collect, process and elaborate video data to quantitatively assess the possible occurrence of conflicts. It integrates computer vision techniques and a conflict detection system to evaluate these interactions by rigorously implementing the theoretical formulation of two primary metrics: Time-to-Collision (TTC) for the pre-event phase and Post Encroachment Time (PET) for the post-event phase. This study is conducted in a real-world setting with mixed traffic conditions to analyse the differences in pedestrian interactions with both human-operated and autonomous vehicles during daytime. The computation of conflict measures allowed us to identify possible conflicts and assess the safety at an unsignalized crossing, in which pedestrians are exposed to more risky conflicts. The results obtained show a higher incidence of more severe conflicts for interactions between pedestrians and human-operated vehicles, which highlights the caution taken in programming the autonomous vehicle.