This study analyzes and compares the determinants of accident severity between human-driven vehicle (HDV) and autonomous vehicle (AV) mixed environments using collision data from the California Department of Motor Vehicles . To address the high dimensionality and categorical complexity of the dataset, an XGBoost-based classification model was developed and the Shapley additive explanations method was employed to explain the contribution and directional influence of each explanatory variable. An undersampling and ensemble approach was utilized to mitigate class imbalances and enhance the model stability. The results revealed that in an HDV environment, driver perception and evasive responses were dominant factors influencing crash outcomes, with collision direction and relative speed significantly affecting the severity. By contrast, in the AV–HDV mixed environment, intersection conditions and complex driving contexts were associated with higher accident severity, thus demonstrating the current limitations of AV systems in managing unstructured traffic scenarios. These findings suggest that as AV deployment progresses, the key determinants of crash severity shift from human behavioral factors to system and environmental factors, thus providing empirical insights for future AV safety evaluations and policy frameworks.

ABSTRACT
1. 서론
2. 관련 문헌 고찰
    2.1. DMV 데이터 기반 자율주행차 사고심각도 연구사례
    2.2. 본 연구의 차별성
3. 연구 설계
    3.1. 변수 정의 및 기술통계 분석
    3.2. 사고 심각도 예측 모델 개발
    3.3. 설명가능한 AI 기법을 활용한 모형 해석
4. 분석 결과
    4.1. 분류 모형 구축 결과
    4.2. SHAP 기반 고위험 요인 비교 분석
5. 결론
감사의 글
REFERENCES

• 박수연(차세대융합기술연구원 첨단교통체계연구실 연구원) | Park Sooyeon
• 가동주(한국도로공사 도로교통연구원 교통연구실 책임연구원) | Ka Dongju
• 김형주(차세대융합기술연구원 첨단교통체계연구실 연구실장) | Kim Hyungjoo Corresponding author