基于驾驶员信心度的SOTIF评价模型建立与试验

基于自动驾驶车辆安全标准ISO/PAS21448《Road vehicles-Safety of the Intended Functionality》(SOTIF),为探究导致SOTIF中人员操作误用的原因,以驾驶员对自动驾驶系统的信任感受为出发点,提出驾驶员对自动驾驶车辆正确操作指令进行错误干预的原因是驾驶员对控制指令没有信心,并引出了用以表现驾驶员与自动驾驶系统之间人机信任感受的评价指标—信心度。为验证信心度指标的实用性,设计了城市道路中常见的转弯和避障工况,并设计了相应的主观问卷,利用层次分析法确定了问卷中不同问题间的权重。利用机器学习中四种分类的方法,建立联系车辆动力学客观参数与驾驶员主观信心感受的主客观评价模型,并选出最优分类器。结果表明,基于马氏距离的分类器准确率最高,可以为自动驾驶车辆控制系统的开发与设计提供支持。

Evaluation Model of SOTIF and Experimental Research Combining Driver Confidence

In order to study the reasons of human misuse in the safety standard ISO/PAS21448“Road vehicles-Safety of the Intended Functionality”(SOTIF)for autonomous vehicles,this paper takes driver’s trust in autonomous driving system as the starting point.The reason why the driver misinterprets the correct command of the autonomous vehicle is that the driver has no confidence in the control command,and this paper puts forward an evaluation indexconfidence to express the human-machine trust between the driver and the autonomous driving system.To verify the practicability of confidence index,common turning and obstacle avoidance conditions in urban roads are designed and corresponding subjective questionnaires are proposed.The weights between five questions in the questionnaire are determined by analytic hierarchy process(AHP).Four classification methods in machine learning are used to establish an evaluation model,combining the objective parameters of vehicle dynamics with the subjective confidence of the driver in the real track test,and the optimal classifier is selected.The results show that the Mahalanobis distance classification method has the highest precision and can support the development and design of the autonomous driving system.