基于交通事故数据的自动紧急制动系统测试场景构建

为了探究面向汽车主动安全技术功能验证的测试场景的科学构建方法，构建符合真实交通状况的高保真测试场景。以自动紧急制动(autonomous emergency braking, AEB)系统为研究对象，以美国高速公路安全管理局事故数据库中筛选出的AEB系统功能适用的6 639起道路交通事故为研究样本，通过机器学习方法实现了由事故数据到测试场景的科学转换。针对传统聚类算法的缺陷，提出了基于层次聚类和K-means聚类相结合的融合聚类算法，并引入聚类曲线以开展事故数据样本的聚类分析。根据聚类获取的12类典型事故场景，构建了面向AEB系统功能验证的14种测试场景。结果表明：相比于传统的K-means聚类算法，融合聚类算法平均减少了8次迭代次数；聚类结果平均减少3%的波动；实现事故数据样本的科学准确聚类且提升数据聚类效率。所提出的测试场景在实现对现有AEB测试场景有效覆盖的同时，为标准测试场景的进一步扩充提供了有力支撑。

Test Scenarios Construction of Automatic Emergency Braking System Based on Traffic Accident Data

In order to explore the scientific construction method of test scenarios for the function verification of vehicle active safety technology, the high-fidelity test scenarios that accord with the real traffic conditions was constructed. Taking the Autonomous Emergency Braking (AEB) system as the research object, and taking the 6 639 road traffic accidents that AEB system functions applicable to which were selected from the accident database of National Highway Traffic Safety Administration as the research samples, the machine learning method was used to realize the scientific transformation from accident data to test scenarios. Aiming at the defects of traditional clustering algorithm, a fusion clustering algorithm based on the combination of hierarchical clustering and K-means clustering was proposed, and the clustering curve was introduced to carry out the clustering analysis of accident data samples. According to the 12 types of typical accident scenarios obtained by the clustering results, the construction of 14 types of test scenarios for the function verification of the AEB system was completed. The results show that the fusion clustering algorithm can reduces 8 iterations on average and the fluctuation of clustering result reduces by 3% on average by comparing with the traditional K-means clustering algorithm. And it realizes the scientific and accurate clustering of accident data samples and improves the efficiency of data clustering. The proposed test scenarios not only effectively cover the existing AEB test scenarios, but also will provide a strong support for the further expansion of the standard test scenarios.