Kang Minhee, Hwang Keeyeon, Yoon Young
School of Electrical Engineering, Korea Advanced Institute of Science and Technology(KAIST), Daejeon, 34141, Republic of Korea.
Department of Computer Engineering, Hongik University, Seoul, 04066, Republic of Korea.
Accid Anal Prev. 2025 Mar;211:107902. doi: 10.1016/j.aap.2024.107902. Epub 2025 Jan 13.
Automated Vehicles (AVs) are on the cusp of commercialization, prompting global governments to organize the forthcoming mobility phase. However, the advancement of technology alone cannot guarantee the successful commercialization of AVs without insights into the accidents on the read roads where Human-driven Vehicles (HV) coexist. To address such an issue, The New Car Assessment Program (NCAP) is currently in progress, and scenario-based approaches have been spotlighted. Scenario approaches offer a unique advantage by evaluating AV driving safety through carefully designed scenarios that reflect various real-world situations. While most scenario studies favor the data-driven approach, the studies have several shortcomings, including perspectives of data, AI models, and scenario standards. Hence, we propose a holistic framework for generating functional, logical, and concrete scenarios. The framework composes explainable scenarios (X-Scenarios) based on real-driving LiDAR data, and visual trend interpretation using eXplainable AI (XAI). The framework consists of four components as follows: (1) voxelization of LiDAR PCD and extraction of kinematic features; (2) classification of critical situations and generation of attention maps using visual XAI and Vision Transformer (ViT) to generate range values of elements in logical scenarios; (3) analysis of the importance and correlations among input data features using SHapley Additive exPlanations (SHAP) for selecting scenarios based on the most relevant criteria; and (4) composition of AV safety assessment scenarios. X-scenarios generated from our framework involve the parameters of ego vehicles and surrounding objects on the highways and urban roads. With our framework highly trustworthy AV safety assessment scenarios can be created. This novel work provides an integrated solution to generate trustworthy scenarios for AV safety assessment by explaining the scenario selection process.
自动驾驶汽车（AVs）正处于商业化的边缘，促使全球各国政府规划即将到来的交通出行阶段。然而，仅靠技术进步并不能保证自动驾驶汽车的成功商业化，还需要深入了解有人驾驶汽车（HV）共存的道路上发生的事故。为了解决这一问题，新车评估计划（NCAP）目前正在推进，基于场景的方法受到了关注。基于场景的方法通过精心设计的场景来评估自动驾驶汽车的驾驶安全性，这些场景反映了各种现实世界的情况，具有独特的优势。虽然大多数场景研究倾向于数据驱动的方法，但这些研究存在一些缺点，包括数据、人工智能模型和场景标准等方面的问题。因此，我们提出了一个用于生成功能化、逻辑化和具体场景的整体框架。该框架基于实际驾驶的激光雷达数据构建可解释场景（X-场景），并使用可解释人工智能（XAI）进行视觉趋势解读。该框架由四个部分组成：（1）激光雷达点云数据的体素化和运动学特征提取；（2）使用视觉XAI和视觉Transformer（ViT）对关键情况进行分类并生成注意力地图，以生成逻辑场景中元素的范围值；（3）使用SHapley Additive exPlanations（SHAP）分析输入数据特征之间的重要性和相关性，以便根据最相关的标准选择场景；（4）构建自动驾驶汽车安全评估场景。我们框架生成的X-场景涉及高速公路和城市道路上的自动驾驶汽车和周围物体的参数。通过我们的框架，可以创建高度可信的自动驾驶汽车安全评估场景。这项新颖的工作通过解释场景选择过程，为生成用于自动驾驶汽车安全评估的可信场景提供了一个综合解决方案。
