Abstract
Busy walking paths, like in a park, city centre, or shopping mall, frequently necessitate collision avoidance behaviour. Lab-based research has shown how different situation- and person-specific factors, typically studied independently, affect avoidance behaviour. What happens in the real world is unclear. Thus, we filmed unscripted pedestrian walking behaviours on a busy urban path. We leveraged deep learning algorithms to identify and extract pedestrian walking trajectories and had unbiased raters characterize situations where two pedestrians approached each other from opposite ends. We found that smaller medial-lateral distance between approaching pedestrians and smaller crowd size predicted an increased likelihood of a subsequent path deviation. Furthermore, we found that whether a pedestrian looked distracted or held, pushed, or pulled an object predicted medial-lateral distance between pedestrians at time of crossing. Our results highlight both similarities and differences with lab-based behaviour and offer insights relevant to developing accurate computational models for realistic pedestrian movement.