Uncovering direct and mediation pathways of autonomous vehicle crash types with augmented built environment factors

Autonomous vehicle (AV) crash types could be critically impacted by diverse contextual characteristics. However, their underlying relationships remain quantitatively underexplored. This study seeks to address this gap by examining the direct and indirect pathways through which contextual variables influence distinct AV crash types.

A total of 584 AV crashes reported to the California Department of Motor Vehicles (CA DMV) between April 1, 2018, and April 12, 2024, were collected for analysis. These data were further enriched with detailed built environment characteristics by geocoding each crash location and retrieving point of interest (POI) information. The AV crash types were regrouped into four categories: rear end, side swipe, angle, and other. Factors encompassing crash, road, temporal, vehicle, and environment characteristics were identified as independent variables. A Multinomial Logistic Regression (MNL) model integrated with a non-parametric bootstrap (1,000 resamples) mediation framework was employed to estimate the direct and indirect effects of significant factors on different AV crash types.

Some variables in direct pathways were captured, such as AV movement preceding collision, vehicle number, VRU, intersection, driving mode, weather, and lighting. Additionally, the mediation pathways were also uncovered, with VRU and intersection emerging as key mediation variables.

These findings contribute to a deeper understanding of AV crash patterns and provide a basis for advancing AV safety technologies.