面向自动驾驶逻辑场景评价的危险域识别方法

Danger Field Identification Method for Evaluation of Autonomous Driving Logic Scenarios

如何评价自动驾驶安全性测试的全面性已成为亟待解决的问题。为此,面向逻辑场景评价提出全面性、准确性、可见性的评价目标。进而,提出面向逻辑场景危险域模态、分布和占比的危险域识别方法,基于离散的具体场景测试结果,在逻辑场景层级全面评估被测系统的安全性。首先,基于Mean Shift算法对具体危险场景聚类实现危险域分区,以发现不同危险场景类别;其次,采用基于特征选择记忆化的决策树算法划分各危险模态的区域边界;最后,解析划分路径以自动化计算危险域占比。为验证所提出的危险域识别方法,将其与基线方法在多模态测试函数上进行对比,发现所提方法在危险域模态、分布和占比方面的计算准确度均优于基线方法。进一步在测试函数和切入逻辑场景中开展应用试验。在测试函数上,基于不同的优化算法进行危险域识别,验证了危险域识别方法的通用性,并基于识别结果对比了不同优化方法的搜索效率。在切入场景上,应用所提方法识别得到切入场景中占比为44%的3个危险域模态及相应的空间分布;并在分析其中2类典型危险场景的基础上,对被测系统提出4项改进要求。研究结果表明:所提出的危险域识别方法可以有效识别自动驾驶逻辑场景中的危险域,在逻辑场景层级全面且准确地评估被测系统安全性。

It is crucial to evaluate the comprehensiveness of safety testing for autonomous driving.Therefore,in this study,evaluation objectives of logical scenario evaluation including comprehensiveness,accuracy,and visibility are proposed.Furthermore,a method based on test results in discrete specific scenarios to identify the danger field in logical scenarios,including danger field modes,distributions,and proportions,is proposed to comprehensively evaluate the safety of the system under test(SUT)at the logical scenario level.First,specific dangerous scenarios were clustered by the Mean Shift algorithm to discover different categories of dangerous scenarios.Second,a Decision Tree by memorization feature selection is proposed to partition the boundary of each danger field mode.Third,the partitioning path was analyzed to automatically calculate the proportion of the danger field.To verify the proposed danger field identification method,it was compared to the baseline on the multimodal test function.The results show that the proposed method is better than the baseline method in the calculation accuracy of the danger field modes,distributions,and proportions.Furthermore,application experiments were conducted on test functions and logical scenario.On the test function,danger field identification was carried out based on different optimization algorithms,verifying the universality of the proposed danger field identification method.The search efficiencies of different optimization methods were compared based on the identification results.In the logical scenarios,three danger field modes with a proportion of 44%and corresponding spatial distributions are identified by the proposed identification method.Based on the analysis of two typical dangerous scenarios,four improvement requirements are proposed for the SUT.The research results show that the proposed method of danger field identification can effectively identify the danger field in logical scenarios of autonomous driving,and comprehensively and accurately evaluate the safety of the SUT at the logical scenario level.