Association rule mining of damage severities in autonomous vehicle rear-end crashes with supplementary built environment data

Abstract

Rear-end crashes represent the most prevalent type of crashes involving autonomous vehicles (AVs), underscoring the critical need to specifically investigate their underlying causes. This study collected 634 AV collision reports from the California Department of Motor Vehicles, spanning from April 1, 2018 to April 12, 2024. An automatic GPU-accelerated variable extraction and enhancement framework was developed to process AV crash records in PDF format. Built environment characteristics, including intersection type, road type, traffic signal, number of lanes, roadside parking, and land use type were supplemented based on extracted geographic coordinates. A total of 2013 human-driven vehicles (HDVs) rear-end crashes that occurred within a 50-foot radius of each AV crash and within the same time window were also collected to conduct the comparative analysis between AV and HDV rear-end crashes. The ordered probit model was employed to identify key individual factors, while the association rule mining was employed to uncover the combinations of contributing factors that frequently co-occur in AV and HDV rear-end crashes. The results confirm the existence of distinct patterns, individual risk factors, and co-occurrence mechanisms that are unique to AV systems under real-world conditions. This study also demonstrates that supplementary built environment factors can significantly influence the occurrence of AV rear-end crashes. Critical factors such as roadside parking, two-way without medians, multiple vehicles, traffic signals, intersections, and land types play pivotal roles in AV rear-end crashes, dominating most rules. These findings can provide valuable insights for reconstructing typical crash scenarios, optimizing road designs, and enhancing AV safety.