面向从城市自然道路交通数据中无监督提取自动驾驶系统测试场景的研究
Toward Unsupervised Test Scenario Extraction for Automated Driving Systems from Urban Naturalistic Road Traffic Data
AI总结：
本文提出无监督机器学习流水线，从城市自然驾驶数据中提取多模态测试场景，并基于 inD 与 Silicon Valley Intersections 比较聚类方法，揭示层次聚类的准确率跃升与饱和规律。
摘要：
基于场景的测试是解决证明配备自动驾驶系统的车辆具有安全行为这一挑战的一种有前景的方法。由于现实世界道路交通中理论上可能出现无限多个具体场景，因此提取与这些系统安全相关行为有关的场景，是其成功验证和确认的关键方面。为此，本文提出一种以无监督方式从自然道路交通数据中提取多模态城市交通场景的方法，以尽量减少（可能存在偏差的）先验专家知识。该方法并非通过将具体场景提取到预定义功能场景中进行（复杂的）基于规则的分配，而是部署了一条无监督机器学习流水线。该方法能够探索数据的未知性质，并将其解释为专家未曾预料到的测试场景。该方法在来自 inD 和 Silicon Valley Intersections 数据集的城市交叉口自然道路交通数据上进行了评估。为此，本文分析了采用哪种聚类方法（K-Means、层次聚类和 DBSCAN）时，场景提取方法表现最佳（参照一个复杂的基于规则的实现）。随后，使用层次聚类的结果表明：当从4个簇增加到5个簇时，总体准确率跃升约20%；而从41个簇开始出现饱和效应，此时总体准确率为84%。这些观察结果可为权衡功能场景数量（即聚类准确率）与测试工作量提供有价值的贡献。本文还讨论了在给定簇总数固定时，观察到的不同簇之间准确率变化的可能原因。
英文摘要：
Scenario-based testing is a promising approach to solve the challenge of proving the safe behavior of vehicles equipped with automated driving systems. Since an infinite number of concrete scenarios can theoretically occur in real-world road traffic, the extraction of scenarios relevant in terms of the safety-related behavior of these systems is a key aspect for their successful verification and validation. Therefore, a method for extracting multimodal urban traffic scenarios from naturalistic road traffic data in an unsupervised manner, minimizing the amount of (potentially biased) prior expert knowledge, is proposed. Rather than an (elaborate) rule-based assignment by extracting concrete scenarios into predefined functional scenarios, the presented method deploys an unsupervised machine learning pipeline. The approach allows exploring the unknown nature of the data and their interpretation as test scenarios that experts could not have anticipated. The method is evaluated for naturalistic road traffic data at urban intersections from the inD and the Silicon Valley Intersections datasets. For this purpose, it is analyzed with which clustering approach (K-Means, hierarchical clustering, and DBSCAN) the scenario extraction method performs best (referring to an elaborate rule-based implementation). Subsequently, using hierarchical clustering the results show both a jump in overall accuracy of around 20% when moving from 4 to 5 clusters and a saturation effect starting at 41 clusters with an overall accuracy of 84%. These observations can be a valuable contribution in the context of the trade-off between the number of functional scenarios (i.e., clustering accuracy) and testing effort. Possible reasons for the observed accuracy variations of different clusters, each with a fixed total number of given clusters, are discussed.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
