基于不同风格行驶模型的自动驾驶仿真测试自演绎场景研究

为了验证自动驾驶汽车决策结果的安全性，提出一种具有自主决策和交互能力的行驶模型生成方法，该行驶模型作为背景车被用于构建自演绎仿真场景来测试自动驾驶汽车的连续决策能力。首先，以强化学习为基础、结合遗传与进化思想，创新地设计并生成了具有自主决策和交互能力的不同风格行驶模型；然后，在模型构建阶段分别训练生成了保守、普通和激进3种风格的行驶模型，其中普通风格行驶模型的训练参数来源于自然驾驶数据集highD的车辆参数分布，保证了该行驶模型的真实性；最后，在普通风格行驶模型的基础上设计并训练出了具有显著激进特征的激进风格行驶模型，以增强自演绎场景的复杂性和测试效果。结果表明：在模型真实性方面，以highD数据集中的跟车速度、车头间距、换道时刻下碰撞时间等参数的分布为真值，研究所生成的普通风格行驶模型的参数分布与真值的平均相似程度为88%，相较于基于规则的智能驾驶人模型（IDM）提升了20.3%；在场景测试性方面，以被测系统为主要责任方的碰撞次数为评估指标，研究生成的不同风格行驶模型所构成的自演绎场景的测试性约是由IDM构成的基线场景的7倍。因此，设计和生成的行驶模型所构成的自演绎场景可以有效支撑面向自动驾驶决策系统的仿真测试。

Self-evolution Scenarios for Simulation Tests of Autonomous Vehicles Based on Different Models of Driving Styles

To verify the safety of the decision-making results of autonomous vehicles (AVs), a method for generating driving models with autonomous decision-making and interaction capabilities was proposed, and the driving models were as background vehicles (BVs) and used to build a self-evolution simulation scenario to test the continuous decision-making capability of AVs. First, based on reinforcement learning and a combination of inheritance and evolution ideas, different driving styles with autonomous decision-making and interaction capabilities were designed in this study. Second, in the model-building stage, three styles of driving models, namely, conservative, general, and aggressive, were generated and trained. The simulation training parameters for the general-style driving model were derived from the parameter distribution of a naturalistic driving dataset named highD to ensure fidelity. Finally, based on this, an aggressive-style driving model with significant aggressive features was designed and trained to enhance the complexity and testing effect of the self-evolution scenario. The results show that the distributions of parameters such as the car-following speed, distance headway, and lane-change moment time-to-collision obtained by using the highD dataset are in agreement with real data. An average similarity of 88% is observed between the general-style driving model generated and the corresponding real data, which is an improvement of 20.3% on the results obtained from the rule-based intelligent driver model (IDM). The proposed self-evolution scenario is seven times more testable than the baseline scenario composed of IDMs for the different driving models generated, as confirmed by the number of collisions in which the system under test is primarily responsible. Thus, self-evolution scenarios composed of the driving models designed and generated in this study can effectively support simulation tests for aiding the decision-making system in AVs.