Classifying crash causation patterns in 2-vehicle collisions between autonomous and conventional vehicles.

Abstract

Objective: This study aims to investigate the causes of 2-vehicle collisions involving an autonomous vehicle (AV) and a conventional vehicle (CV). Prior research has primarily focused on the causes of crashes from the perspective of AVs, often neglecting the interactions with CVs.

Method: To address this limitation, the study proposes a classification framework for crash causation patterns in 2-vehicle collisions involving an AV and a CV, considering their interactions. The framework categorizes the crash causation patterns into 5 distinct types: (1) failure of the AV system, (2) failure of takeover control, (3) driver error after takeover, (4) CV failure to adapt to unforeseen changes in AV behaviors, and (5) other factors related to the CV. Utilizing the AV crash data set proposed by Zheng et al., this study extracted 450 two-vehicle collisions involving AVs and CVs for our analysis.

Results: Our analysis reveals that the majority of 2-vehicle collisions are triggered by CVs, specifically identified in patterns 4 and 5. Pattern 4 is the primary crash causation factor, accounting for 55% of total collisions. The leading contributing factor to pattern 4 is the improper response of CVs to AVs stopping. There are notable variations in crash injury severity, collision type, and environmental conditions across different causation patterns. Crashes stemming from human drivers' errors (patterns 3, 4, and 5) are more likely to result in moderate to severe injuries. Specifically, pattern 4 notably exhibits the highest likelihood of causing rear-end collisions, whereas patterns 1 and 5 are more prone to causing side collisions. Additionally, crashes associated with pattern 4 are more frequently observed in locations with traffic controls or obstacles.

Conclusion: Based on our findings, we propose several recommendations for manufacturers to enhance AV safety performance. These include minimizing planning errors in autonomous driving algorithms, improving communication abilities between AVs and other road users for smoother interactions and better anticipation of actions, and providing specialized driver training for navigating mixed environments with both AVs and CVs.