911³Ô¹Ï

Predicting Pedestrian Collision Avoidance Behaviours: Insights from Real-World Footage and Virtual Reality Interactions

This seminar investigates collision avoidance behaviours among pedestrians in real-world and virtual environments, aiming to enhance computational models and improve urban safety. While laboratory studies typically study individual factors influencing pedestrian behaviour in isolation, real-life interactions involve a complex interplay of spatial, situational, and person-specific elements. To bridge this gap, I conducted three experiments. In one experiment, I captured unscripted one-versus-one pedestrian interactions along a busy urban path in Metro Vancouver. I analyzed videos with deep learning algorithms to extract pedestrian trajectories and had unbiased raters characterize interactions. Results revealed that reduced medial-lateral (ML) separation between approaching pedestrians increased the probability of path deviation, while larger crowd sizes reduced this likelihood. Furthermore, distraction significantly increased ML separation at the time of crossing, whereas mobility constraints significantly reduced it. Overall, these results suggest that pedestrians actively manage a personal space envelope. In a second experiment, I translated these insights into a controlled virtual reality (VR) environment that replicated the urban setting. Participants walked along a 3.5-m-wide virtual path and encountered a virtual pedestrian with a mobility constraint (i.e., pushing a stroller, carrying a shopping bag, or pushing a bicycle) combined with different distraction levels. Results showed that when the virtual pedestrian pushed a stroller, the participants deviated earlier and maintained larger ML separation at the time of crossing. In contrast, distraction had no significant effect on avoidance behaviours. In a third experiment, I examined unscripted one-versus-group pedestrian interactions using the videos from the same location and time as in Experiment 1.  Results revealed that reduced initial ML separation and crowd size predicted a higher likelihood of deviation, while greater crowd size and distractedness increased ML separation at crossing. In addition, larger groups were associated with reduced ML separation compared to smaller groups, underscoring the influence of group dynamics on avoidance strategies. Together, these findings demonstrate that collision avoidance in natural pedestrian environments is governed by multiple simultaneous factors. The results provide empirical data to help refine pedestrian simulation models, and they may inform urban planning and guide the development of assistive navigation systems and realistic VR simulations.