面向自动驾驶仿真测试的高覆盖切入场景库生成方法  

High-coverage Cut-in Scenario Library Generation for Automated Driving Simulation Testing

作　　者：李鹏辉 董倩茹 袁赫男 胡文浩 孙巍 谷远利[1] 董春娇[1] LI Peng-hui;DONG Qian-ru;YUAN He-nan;HU Wen-hao;SUN Wei;GU Yuan-li;DONG Chun-jiao(School of Traffic and Transportation,Beijing Jiaotong University,Beijing 100044,China;State Key Laboratory of Intelligent Vehicle Safety Technology,China Automotive Engineering Research Institute Co.Ltd.,Chongqing 401122,China;State Key Laboratory of Market Regulation(Defective Product and Safety),SAMR Defective Product Recall Technical Center,Beijing 100101,China;Key Laboratory of Ministry of Public Security for Road Traffic Safety,Traffic Management Research Institute of the Ministry of Public Security,Wuxi 214151,Jiangsu,China)

机构地区：[1]北京交通大学交通运输学院,北京100044 [2]中国汽车工程研究院股份有限公司智能汽车安全技术全国重点实验室,重庆401122 [3]国家市场监督管理总局缺陷产品召回技术中心国家市场监管重点实验室(产品缺陷与安全),北京100101 [4]公安部交通管理科学研究所道路交通安全公安部重点实验室,江苏无锡214151

出　　处：《中国公路学报》2024年第7期237-249,共13页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(52372422,52302425);道路交通安全公安部重点实验室开放课题基金项目(2023ZDSYSKFKT07)。

摘　　要：为解决虚拟仿真测试所需高覆盖场景的生成难题,以自动驾驶车辆事故高发的高速公路切入场景为研究对象,基于搭载多源传感器实车采集的自然驾驶数据,开发基于规则的切入场景自动提取算法;选取核心场景要素建立基于运动学特征的车辆切入轨迹模型,量化分析模型参数分布特征;基于高斯混合模型构建多维度逻辑场景参数联合概率密度函数,在此基础上提出基于哈密尔顿蒙特卡洛(Hamiltonian Monte Carlo,HMC)采样与Jensen-Shannon(JS)散度覆盖度表征的多维空间场景参数高覆盖生成方法。基于所提取的2422例车辆切入片段研究发现:①基于起始时刻主车速度Ve0、相对速度Vr0、车距Dx0、切入时长T、切入车辆横向加速度ay和纵向加速度ax六参数的横纵向运动学模型可有效表征切入车辆运动轨迹,平均拟合均方根误差为0.7 m;②八分量高斯混合模型对切入场景参数的联合概率密度分布拟合效果最佳;③JS散度随着场景采样数量的增加快速下降而后逐渐收敛至0.01,表明HMC方法可实现切入场景参数的快速采样与高覆盖生成;④本方法实现切入数据集片段信息全覆盖所需场景生成数量为2160个,相比于传统马尔科夫蒙特卡洛(Markov Chain Monte Carlo,MCMC)方法所需场景生成量缩小约73倍,测试效率显著提高,推荐用于高速公路全量切入场景库构建。提出的切入场景轨迹模型与高覆盖生成算法具有可解释性、高覆盖度、生成快捷的特点,将为自动驾驶虚拟仿真测试提供有力支撑。To address the challenge of generating high-coverage scenarios for virtual simulation testing,this study focused on highway cut-in scenarios,which are prone to accidents of automated vehicles.Based on naturalistic driving data collected by vehicles equipped with multiple sensors,a rule-based algorithm was developed to extract cut-in scenarios automatically.Then,the core scenario elements were selected to establish a vehicle cut-in trajectory model based on kinematic features.Next,we analyzed the distribution characteristics of the model parameters quantitatively.Subsequently,a multi-dimensional logical scenario parameter joint probability density function was constructed using a Gaussian mixture model(GMM).Based on this,a high-coverage generation method for multi-dimensional spatial scenario parameters is proposed.It is based on Hamiltonian Monte Carlo(HMC)sampling and Jensen-Shannon(JS)divergence verification.Based on 2422 vehicle cut-in segments,we found that①a six-parameter kinematic model effectively represents the cut-in vehicle trajectory,with a root mean square error of 0.7 m;②an eight-component GMM provides the best fit for the joint probability density distribution of cut-in scenario parameters;③the JS divergence decreases rapidly as the number of scenario samples increases and then gradually converges to 0.01;this suggests that the HMC method facilitates rapid sampling and high-coverage generation of scenario parameters;and④the number of scenarios required by the proposed method to achieve full coverage of the dataset segment information is 2160.Compared to traditional Markov Chain Monte Carlo methods,the required number of scenarios to be generated is approximately reduced by a factor of 73,significantly improving the testing efficiency.The proposed approach is recommended for constructing a full-highway cut-in scenario library.The proposed scenario trajectory model and high-coverage generation algorithm have the characteristics of interpretability,high coverage,and rapid generation,and will pr

关 键 词：交通工程 高覆盖场景生成 哈密尔顿蒙特卡洛 仿真测试场景 自动驾驶 切入 

分 类 号：U467.1[机械工程—车辆工程]