自动驾驶系统在复杂、开放、不确定的真实道路环境中依然面临巨大的安全风险挑战。本文以美国加州机动车车辆管理局公开的710条自动驾驶车辆事故数据为基础，考虑自动驾驶车辆感知-决策-执行过程任务的系统特性，从人车交互、道路条件、环境因素等四个维度出发构建事故场景画像指标体系，综合运用多重对应分析揭示事故多维度变量间的内在关联，建立了基于SOM拓扑神经网络与K-means聚类算法的典型场景画像方法，识别高、中、低风险要素组合。通过数据挖掘技术深入分析自动驾驶车辆与静止障碍物、行人/非机动车、运动车辆碰撞事故的特征及影响因素，进而构建事故典型场景画像。数据表明，事故发生时73.94%的车辆投入使用时间不足两年，没有物理分隔的道路上发生的事故占比77%，在交叉口发生的事故占比50%以上。典型场景分析发现，三类事故的主导因素与场景特征存在相同点和不同点：首先，自动驾驶车辆与三种对象碰撞场景的重要影响因素均包含决策算法（规则驱动和双驱动）和感知系统（视觉主导和雷达主导）。但是，与静止障碍物碰撞的典型场景因素涉及光照条件和人类驾驶模式；与行人/非机动车碰撞常关联于自动驾驶车辆的控制架构类型；而与运动车辆的碰撞则主要涉及所处区域环境。研究精准刻画三类典型场景的精细化画像，为自动驾驶系统的测试优化及安全风险评估提供了关键的数据支撑与理论依据，助力提升自动驾驶技术的整体安全水平。

Autonomous Driving Systems continue to face significant safety challenges in complex, open, and uncertain real-world road environments. Based on 710 publicly available autonomous vehicle accident reports from the California Department of Motor Vehicles, this study conducts an in-depth analysis of the characteristics and influencing factors of three types of ADS-involved collisions: those with static obstacles, pedestrians/non-motorized vehicles, and moving vehicles, with the aim of constructing typical scenario profiles for these accidents. Taking into account the inherent characteristics of ADS in performing perception-decision-execution tasks, an accident scenario profiling index system is developed across four dimensions： human-vehicle interaction, system characteristics, road conditions, and environmental factors. The research comprehensively employs Multiple Correspondence Analysis to uncover intrinsic relationships among multi-dimensional qualitative variables in accidents. Furthermore, it establishes a typical scenario profiling methodology by integrating Self-Organizing Map (SOM) neural networks with the K-means clustering algorithm to identify high, medium, and low risk factor combinations. Through data mining techniques, the characteristics and influencing factors of collisions involving autonomous vehicles with static obstacles, pedestrians/non-motorized vehicles, and moving vehicles are deeply analyzed, leading to the construction of typical accident scenario profiles. The data indicates that 73.94% of the vehicles involved in accidents had been in service for less than two years, 77% of the accidents occurred on roads without physical separations, and over 50% took place at intersections. The analysis of typical scenarios reveals both similarities and differences in the dominant factors and scenario characteristics across the three accident types. Firstly, significant influencing factors common to collisions with all three types of objects include the decision-making algorithm (rule-driven and dual-driven) and the perception system (vision-dominant and radar-dominant). However, typical scenarios for collisions with static obstacles involve lighting conditions and human driving modes; those with pedestrians/non-motorized vehicles are often related to the type of autonomous vehicle control architecture; while collisions with moving vehicles primarily involve the regional environment. Ultimately, the study precisely delineates refined profiles for the three types of typical scenarios, providing crucial data support and a theoretical basis for the testing, optimization, and safety risk assessment of Autonomous Driving Systems, thereby contributing to the enhancement of the overall safety level of autonomous driving technology.