Causes and Effects of Autonomous Vehicle Field Test Crashes and Disengagements Using Exploratory Factor Analysis, Binary Logistic Regression, and Decision Trees

Abstract

Autonomous vehicles (AV) are being widely tested around the world. California is one of the jurisdictions permitting extensive AV testing. Field testing resulted in approximately 0.02 AV crashes per 1,000 AV miles traveled, as well as incidents leading to driver disengagement of the AV systems. Factors related to human error, system failure, surrounding vehicles, and roadway features could cause an AV-involved crash to occur or result in pre-crash disengagement. This study focuses on AV crashes in which the AV was operating in active AV mode or shortly after the test operator had disengaged the AV to resume conventional control. AV crash data were extracted for the years 2017 to 2021 from California’s AV crash database maintained by the California Department of Motor Vehicles (DMV) to account for the rapid pace of AV development and focus on current causes of incidents. This paper utilized multiple statistical approaches to quantitatively investigate AV crashes and disengagement events. Investigation of latent manifest using exploratory factor analysis (EFA) was conducted and ordinal logit models and decision trees were employed in this study. EFA clarified latent variables that could identify an AV crash, in which operator involvement, incorrect maneuver decision, crash severity, and environmental conditions were the manifests obtained from the analysis. Results from the logistic regression and the decision trees showed that collision type, AV movement type, and other vehicle movement type are significant factors contributing to AV crashes.