Guo Rui, Chen Yanyan, Zhang Yunchao, Yi Fuhua
Beijing Key Laboratory of Traffic Engineering, Beijing University of Technology, Beijing, China.
Beijing Lingyun Technology Co., Ltd, Beijing, China.
Traffic Inj Prev. 2025 Jul 16:1-8. doi: 10.1080/15389588.2025.2520006.
The market for autonomous vehicles (AVs) is developing rapidly, while safety concerns persist as a critical challenge hindering their widespread development and commercialization. Intersections, characterized by highly dynamic and unpredictable traffic conditions, represent particularly high-risk scenarios for AVs. This study systematically investigates the key risk factors influencing collision severity at intersections to enhance AV safety performance.
This study employs the Autonomous Vehicle Operation Crash Dataset Across The Globe (AVOID) to investigate the key factors influencing the injury severity of AV related collisions. A hybrid analytical framework is proposed, integrating an XGBoost-SHAP model for feature importance analysis and a multinomial logit (MNL) model for statistical inference. Following the factor importance ranking, nine critical determinants were selected to examine their individual effects on injury severity levels. Furthermore, the XGBoost-SHAP approach was utilized to explore interaction effects among significant factors, revealing synergistic relationships between key features.
The results indicate that the majority of crashes at intersections occurred when AVs were stationary or moving at low speeds (17.71% while stopped, 42.76% at speeds below 10 mph). Approximately 58.14% of the crashes involved autonomous driving mode, with an injury rate 14.5% higher compared to manual mode. Factors such as pre-crash movement, crash scene, contact area, pre-crash speed, and autonomous mode significantly influenced injury severity. Crashes occurring during straight-line travel or lane changes in autonomous mode tended to result in more severe injury compared to manual driving. Additionally, crashes in steering direction scenes at speeds between 10 and 20 mph were associated with higher injury severity, and speeds exceeding 20 mph in traffic through scenes led to even more severe injuries.
This study reveals the main factors influencing the severity of collisions in autonomous vehicles and the combination of the following factors that may increase the severity of injuries: autonomous driving mode, lane changing, turning or crossing scenarios, and high-speed driving. The results advance the understanding of autonomous vehicle safety and offer potential implications for enhancing self-driving systems.
自动驾驶汽车（AV）市场正在迅速发展，而安全问题仍然是阻碍其广泛发展和商业化的关键挑战。交叉路口具有高度动态和不可预测的交通状况，对自动驾驶汽车来说是特别高风险的场景。本研究系统地调查了影响交叉路口碰撞严重程度的关键风险因素，以提高自动驾驶汽车的安全性能。
本研究采用全球自动驾驶汽车运行碰撞数据集（AVOID）来调查影响自动驾驶汽车相关碰撞伤害严重程度的关键因素。提出了一个混合分析框架，集成了用于特征重要性分析的XGBoost-SHAP模型和用于统计推断的多项逻辑回归（MNL）模型。根据因素重要性排名，选择了九个关键决定因素来检验它们对伤害严重程度水平的个体影响。此外，利用XGBoost-SHAP方法探索显著因素之间的相互作用效应，揭示关键特征之间的协同关系。
结果表明，交叉路口的大多数碰撞发生在自动驾驶汽车静止或低速行驶时（停车时为17.71%，速度低于10英里/小时时为42.76%）。大约58.14%的碰撞涉及自动驾驶模式，与手动模式相比，受伤率高14.5%。碰撞前的运动、碰撞场景、接触面积、碰撞前速度和自动驾驶模式等因素显著影响伤害严重程度。与手动驾驶相比，在自动驾驶模式下直线行驶或变道时发生的碰撞往往导致更严重的伤害。此外，在10至20英里/小时速度下转向方向场景中的碰撞与更高的伤害严重程度相关，在交通通过场景中速度超过20英里/小时会导致更严重的伤害。
本研究揭示了影响自动驾驶汽车碰撞严重程度的主要因素以及可能增加伤害严重程度的以下因素的组合：自动驾驶模式、变道、转弯或交叉场景以及高速行驶。研究结果推进了对自动驾驶汽车安全性的理解，并为增强自动驾驶系统提供了潜在的启示。
