Abstract: Reasonably foreseeable scenario generation is vital for safety verification of autonomous driving systems. To tackle this problem, this study proposes a framework to generate data-driven simulation scenarios. Firstly, data tags are derived from naturalistic driving data set and categorized into four categories: road topology, ego car behaviors, obstacle behaviors and driving environment; Then, logic operations defined upon data tags are proposed to describe formal and machine-readable abstract scenarios. The cut-in scenario is selected to illustrate logical scenario generation, due to its high frequency occurrence and potential safety impact on road. Specifically, kinematic parameters such as the relative distance and speed between the ego car and the interactive obstacle are used to describe the cut-in scenario. Based on real road test data, linear regression model is applied to fitting parameter space boundaries. Finally, procedure of concrete safety scenario generation is proposed on the basis of the derived parameter space. Experiments show that simulation scenarios generated following the proposed framework have consistent distributions with real-world natural driving scenarios with respect to critical parameters, and are more safety-critical, thus enabling more efficient safety evaluation for autonomous driving system.

Key words: autonomous driving; simulation scenario generation; scenario tags; event expression; parameter space analysis

摘要： 合理可预见的仿真测试场景生成对于自动驾驶系统的安全性验证至关重要。为了解决 该问题，文章提出了一个完整的数据驱动的仿真安全场景生成流程与框架。该框架首先基于 自然驾驶数据提取出道路拓扑、主车行为、障碍物行为、驾驶环境等数据标签；然后，定义 事件表达式对数据标签进行逻辑组合，构建形式化的抽象场景；针对实际道路环境中高频发 生且存在潜在安全影响的切入类场景，定义主车与障碍物之间的相对位置、速度等运动学参 数描述切入逻辑场景。基于真实的自然驾驶数据集，运用线性回归的参数空间拟合方法，给 出了切入场景的参数空间边界，以及生成合理可预见的安全测试场景的方法。最后，设计实 验证明了依据该方法生成的场景，不仅能与真实的自然驾驶场景在关键参数的分布上保持一 致性，同时具有更强的安全挑战性，能更有效地评估自动驾驶系统的安全边界。

关键词: 自动驾驶；仿真场景生成；场景标签；事件表达式；参数空间分析