Quantifying risk factor influences in autonomous vehicle collisions: a Bayesian network probabilistic analysis

The frequent occurrence of Autonomous Vehicle (AV) collisions significantly impacts development and user trust. These collisions arise from a complex interaction of factors, but their interdependencies are not fully understood. This study analyzed 776 publicly available AV -related collision reports from the California Department of Motor Vehicles, identifying key factors and their complex interactions. The identified risk factors are divided into three categories: vehicle basic information, collision details, and road and environmental characteristics.

The statistical analysis method of Chi-square test was used to evaluate the significance of single factors. Bayesian network analysis further exploratorily constructs causal chains and examines the impact of each factor on collision severity.

Six variables, including vehicle mode, brand, collision type, pre-collision motion, weekday and roadway type, can independently affect the collision severity. Bayesian network exploration found that brand affects vehicle mode, vehicle mode affects pre-collision motion, and pre-collision motion affects collision type. Side swipe collisions, rear-end collisions, road sections, stationary or slow-moving conditions are the most likely to cause property damage. Casualties are most likely to occur in incidents involving broadside collisions and highways. Additionally, intersections are high-risk collision locations. The autonomous driving mode is similar to the conventional human driving mode in terms of collision risks, but there are still certain safety hazards, such as a higher probability of a broadside collision.

These findings show that AV technology should be continuously improved in many aspects such as environmental perception, decision-making algorithms, and safety mechanism design to improve the overall safety and reliability of AV and make it better integrate into daily life.