Analysis of Autonomous Vehicle Crash Mechanisms: Exploring with Machine Learning and Association Rules

Abstract

This study investigates the impact of technological innovations and regulatory updates on the safety outcomes of autonomous vehicles (AVs), aiming to identify which machine learning algorithms yield the most reliable results in analyzing AV crash data. With advancements in robotics and artificial intelligence, AVs have the potential to significantly reduce damages from driving crashes by eliminating human error, thereby enhancing safety. Utilizing the synthetic minority over-sampling technique (SMOTE) for data balancing, the research assesses its effects on predictive accuracy. This study addresses safety gaps by examining crash mechanisms related to AVs using machine learning methods and association rules such as Apriori and classification and regression trees (CART), exploring limitations in current research, particularly the lack of AV crash data, the relationship between land use and crash severity, and interactions involving driverless vehicles. To bridge these gaps, a dataset of 606 reported AV crashes from the California Department of Motor Vehicles (CA DMV) is analyzed to evaluate the current safety status of AVs and identify essential factors for acceptable safety performance levels in the future. Results show that 15.4% of crashes led to bodily injury, with intersections featuring traffic lights identified as primary locations for severe crashes, especially affecting vulnerable users like cyclists. Most driverless vehicle collisions and AV-to-AV incidents were classified as sideswipe collisions. While driverless vehicles reduce rear-end and broadside collisions, their interactions with human-driven vehicles present unique challenges that necessitate technological innovations, the development of vehicle-to-vehicle (V2V) communication, and legal updates. This study provides valuable insights for planners, transportation engineers, and researchers to enhance mixed traffic safety.