Trajectory-based conflict investigations involving two-wheelers and cars at non-signalized intersections with computer vision

Unsafe acts at non-signalized intersections have become a primary contributor to traffic accidents and fatalities. Whereas many studies have focused on non-signalized intersections in the past decade, capturing and recording road users' micro-behavior and risk in the mixed traffic flow remain challenging. A large number of two-wheelers (e.g., bicycles and e-bikes) appear at non-signalized intersections, in which the conflict behaviors are highly unpredictable. In this study, conflicts involving two-wheelers and cars at non-signalized intersections were investigated based on trajectory data collected with a YOLOv3-based framework automatically. A novel conflict identification algorithm was developed to gather and process microscopic trajectory data. To detect conflict behaviors involving two-wheelers and cars, near-crash identification was employed with a post-encroachment time indicator that also contributes to demonstrating the effect of vehicle order on conflict severity from an unprecedented perspective. The proposed framework was applied to a case study at a university campus in Shanghai. To explore the relationship between contributing factors and conflict severity, a significance test and ordered probability models were implemented using 10,304 conflicts collected from video data. The statistical analysis disclosed that conflicts involving e-bikes accounted for the highest proportion, and the order of vehicles (e.g., pre-encroachment vehicle and post-encroachment vehicle) has different effects on conflict severity. The analytical results with risk assessment can contribute to developing intersection-specific countermeasures for traffic safety from the perspectives of education, engineering, and law enforcement. The trajectory-based framework can be adapted to intelligent transportation systems to enhance safety management.