Abstract: Objective Given the limitations of conventional methods for the quantitative assessment of the perception capability of advanced driver assistance systems (ADAS), an investigation and analysis technology was established to retrospectively evaluate the object recognition functional performance prior to a crash, providing technical support for accident causation determination and ADAS-related function investigation. Methods On-board video from the accident vehicle was used as the primary data source. A YOLOv8 object detection model was trained using the publicly available KITTI 2D dataset and a traffic cone dataset. Frame-by-frame detection was performed, and the detection time, model output category, and temporal changes of key objects were extracted and cross-compared with the ADAS perception data recorded in the accident vehicle’s operation log. Results The YOLOv8 model detected the vehicles, the road workers, and the traffic cones in the accident vehicle’s on-board video, with the first detection frame of the vehicle object approximately 4.97 s from the collision frame. At each sampling time within 7 s before the collision, the accident vehicle’s operation log recorded the object recognition type as “no object”. Comparison of the two sets of data showed differences between the external video detection results and the object recognition records in the log data, and the accident vehicle missed relevant objects in the actual scenario, indicating that the accident vehicle’s perception system may have deficiencies in the perception and recognition of critical objects, including stationary vehicles, traffic cones, and road workers. Conclusion The method proposed in this study can reconstruct the detection history of key objects in accident videos and analyze differences in ADAS object recognition before the accident, enabling retrospective analysis of the object recognition function of the accident vehicle’s perception system. It provides technical support for accident investigation involving assisted-driving vehicles and for the improvement of ADAS perception performance.

Key words: operational data, object detection, YOLOv8, assisted driving, perception system, accident investigation

摘要： 目的 针对辅助驾驶汽车事故调查中传统方法难以量化高级驾驶辅助系统（advanced driver assistance systems，ADAS）感知能力的问题，提出一种可追溯事故前目标识别功能表现的调查分析技术，为事故致因分析和ADAS相关功能调查提供技术支撑。方法 以事故车辆车载视频为核心数据源，采用KITTI 2D公开数据集与交通锥数据集开展YOLOv8目标检测模型训练，对实际事故车载视频进行逐帧检测，提取关键目标的检出时刻、模型输出类别及时序变化，并与事故车辆运行日志中记录的ADAS感知数据进行交叉比对。结果 YOLOv8模型在事故车辆车载视频中检出了车辆、作业人员及交通锥等目标，其中，车辆目标首次检出帧距碰撞帧约4.97 s。事故车辆运行日志在碰撞前7 s的各采样时刻，目标识别类型均记录为“无目标”。通过两组数据对比分析，外部视频检测结果与日志数据的目标识别记录存在差异，事故车辆在实际场景下对相关目标存在漏检，表明事故车辆感知系统对静止车辆、交通锥及作业人员等关键目标的感知识别可能存在不足。结论 本研究所提出的技术能够重构事故视频中关键目标的检出历程，可分析ADAS在事故发生前的目标识别差异，实现对事故车辆感知系统目标识别功能的追溯分析，为辅助驾驶汽车事故调查及ADAS感知性能改进提供技术支持。

关键词: 运行数据, 目标检测, YOLOv8, 辅助驾驶, 感知系统, 事故调查