基于驾驶负荷的人机共驾安全性评价方法研究

The advancement of autonomous driving will most likely be a gradual progress and thus the human driver will be kept as a co-pilot of the vehicle for a great while. This co-piloting way of driving will almost certainly have significant impacts on driving safety. Driver's workload, whose relevance to driving safety has been proved to be high, shall be used with high accuracy and confidence, to evaluate the safety when vehicles are co-piloted by human and autonomous driving systems...In order to achieve this goal, new influencing factors introduced by co-piloting and their influencing mechanism on driving workload will need to be identified and studied at the quantitative level. A quantitative model should then be constructed to describe precisely how driving workload would change as its factors change. The proposed research is going to be conducted accordingly...Firstly, parameters that characterize driving workload will be extracted from multiple-resources information, including drivers' physiological parameters, records of drivers' operations on the vehicle , and traces of vehicles' motion. Approaches such as "analysis of variance" and "Principal components analysis (PCA)" will be adopted to screen and afterwards create the group of evaluation indicators of driving workload with a reasonable level of complexity and a high level of accuracy. Driving workload will be concluded after two stages of fusion of such indicators, and a comprehensive quantitative evaluation model of driving workload with high accuracy as well as robustness is then constructed...Secondly, various classes of key influencing factors of driver mental workload introduced by co-piloting will be analyzed, including the complexity of the driving environment, the level of automation, the automated driving mode, and the mechanism adopted to transit control between human drivers and autonomous driving systems. Quantifiable factors will be identified with experts' advices as inputs. With orthogonal experiments conducted on a driving simulator platform, factorial analysis will be done to obtain the matrix of weighting coefficients of all factors. On top of all that, a predictive model of driving workload is constructed and validated...Lastly, on basis of the relationship between driving workload and driving safety described in the inverted-U model, driving safety can be classified into different levels with respect to classifying thresholds of the quantified driving workload.

在自动驾驶发展过程中，人机协同驾驶(人机共驾)的现象将长期存在并对驾驶安全性产生显著影响。驾驶安全性与驾驶负荷密切相关，因此从定量分析的角度深入研究人机共驾模式下驾驶负荷的影响因素及影响机理，揭示其变化规律并建模描述，是准确评价人机共驾安全性的有效方法。本项目基于驾驶仿真和实车实验，从驾驶人生理信号、车辆操纵及运动行为等多源信息中选取驾驶负荷表征参数，利用方差分析、主成分分析等方法筛选并构 建评价指标体系，通过指标的二级融合构建准确性高、鲁棒性强的驾驶负荷定量评价模型;分析人机共驾阶段驾驶环境、自动驾驶等级及驾驶任务类型、驾驶控制权切换机制等关键影响因素，基于正交实验研究上述因素对驾驶负荷的影响机理，建立可量化的因素组并确定各因素的权重系数矩阵，构建基于多因素的驾驶负荷预测模型并验证;最终基于驾驶负荷与驾驶安全性的倒U 形关系，结合分类阈值，实现基于驾驶负荷的人机共驾安全性评价。

当前，L4级以上高等级自动驾驶虽然在特定商业模式下开始尝试产业化落地，但是其应用场景极为有限，其安全性也远未得到充分验证。全域高等级自动驾驶的落地仍然面临突出的技术瓶颈和社会困境，短期内难以突破。驾驶人和自动驾驶协同完成驾驶任务的人机共驾模式成为实现智能网联汽车从低等级向高等级自动驾驶平稳过渡的重要途径。.在人机共驾模式下，驾驶控制权在驾驶员与自动驾驶系统两个控制实体之间频繁转移，驾驶人的行为特性由此发生显著变化，并深刻影响驾驶安全性。因此，对人机共驾安全性进行有效评价是自动驾驶发展及过程中所需要必须解决的关键问题。.本项目针对人机共驾的安全性评价问题，提出以驾驶负荷为评价对象，建立人机共驾安全性的有效评价方法。项目首先构建了基于多模态生理参数、车辆驾驶绩效、驾驶人监控视频等多源信息融合的驾驶负荷定量综合评价模型，实现了车载环境下的驾驶负荷鲁棒、可信量化评估。以此为基础，项目研究了驾驶环境复杂度、自动驾驶等级及自动驾驶任务类型、驾驶控制权切换机制等多因素耦合影响下的驾驶负荷动态演变机理。为了定量分析复杂驾驶环境对驾驶负荷的影响，提出了表征驾驶环境复杂度的指标体系及量化方法，完成了针对高速公路和城市道路的环境复杂度量化评价模型构建。进一步地，项目构建了考虑“人-车-环境”因素的面向行驶环境全域的人机共驾机制，建立了基于行驶风险量化评估的人机驾驶控制权权重动态分配策略，并基于间接式共享控制框架，建立了人机共驾控制策略。最后，在完成驾驶负荷定量评价和驾驶负荷影响机理辨识的基础上，建立了驾驶负荷预测模型，结合驾驶负荷与驾驶安全性的倒U形关系，实现了基于驾驶负荷的人机共驾安全性有效、可信评价。项目研究成果为人机共驾机制构建及其安全性测试评价提供了方法借鉴和模型基础，可有效地从测试评价的角度驱动人机共驾型自动驾驶系统开发，推动人机共驾相关研究从理论模型向实车应用加速转变。