Related Experiment Video
Updated: Jul 21, 2026
Tactile Vibrating Toolkit and Driving Simulation Platform for Driving-Related Research
Published on: December 18, 2020
Improving our understanding of contributing factors in California level 4 autonomous vehicle crashes through latent
Jiacheng Wang1, Corey D Harper2, Chris Hendrickson1
1Carnegie Mellon University, Department of Civil and Environmental Engineering, 4815 Frew St, Pittsburgh PA, US 15213, United States.
Introduction:
Autonomous vehicles (AVs) could lower crash frequency and injury severity by reducing the number of crashes caused by human-error. Using the National Highway Traffic Safety Administration Standing General Order database-we revisit California AV crashes to determine: (i) which factors drive injuries; (ii) whether their effects are uniform or scenario-dependent; and (iii) how pairs of factors interact to raise or lower injury risk.
Method:
We examined 947 level 4 AV crashes. Latent class analysis grouped crashes into homogeneous scenarios. Within each scenario we trained a machine learning classifier, balanced with random oversampling, to predict injury versus no-injury outcomes. Using SHapley Additive exPlanations (SHAP) the main effects of different variables and their pairwise interaction effects were analyzed.
Results:
We identified three latent crash profiles: low-speed daytime, moderate-speed across different times, and low-speed nighttime. Rear-end collisions and turning maneuvers are positively correlated with injury outcome in both low speed profiles and have context-dependent impact in the moderate-speed profile. Non-motorist crashes carry elevated injury risk across all classes, with the highest SHAP values observed in the moderate-speed profile. Adverse weather increases injury likelihood in the low-speed daytime profile on streets, while wet or icy roadway surfaces raise injury likelihood in the moderate-speed crash profile. Time-of-day interaction effects are observed across several scenarios.
Conclusions:
Our LCA-Machine Learning-SHAP framework reveals key factors and interaction effects on level 4 AV crash injuries in an interpretable way. Further insights into contributing factors can be assessed as larger datasets become available.
Practical Applications:
Improving predictive driving algorithms so that AVs could better anticipate changes in vehicle speed, acceleration, and braking, limiting speeds in areas with high pedestrian activity, deploying friction-aware controls to improve traction on non-dry roads, and enhancing weather/visibility classification for AVs to automatically adjust speed and headway in adverse weather conditions.
