人机共驾智能车驾驶模式决策属性析取研究

为了深入分析驾驶模式决策影响因子，通过实车试验采集了人-车-路多源特征信息。用驾驶人主观经验将驾驶模式划分为人工驾驶、警示辅助、自动驾驶3种状态，并利用采集的驾驶人血流量脉冲（BVP）和皮肤电导（SC）值进行K均值聚类，将驾驶人当前合适的驾驶模式自动聚类为3级。通过融合驾驶人自汇报结果和聚类结果对驾驶模式进行准确标定。采用以信息增益为依据的Ranker算法对多特征进行排序，并在此基础上，根据多分类器分级结果确定最优特征属性集合。研究结果表明：当选取车速、车头时距、车道中心距离、前轮转角标准差、驾驶经验5个指标为特征子集时，支持向量机、朴素贝叶斯及K近邻这3种分类器的识别准确率都超过90%；除警示辅助模式与自动驾驶模式下的车速值和车道中心距之外，其余所有不同模式决策属性值均呈显著性差异；研究结果可为人机共驾智能车驾驶模式决策提供依据。

Research on Impact Factors Extraction for Driving Mode of Intelligent Vehicle

In order to explore the significant impact factors of intelligent driving mode selection, the multi-source feature information from driver-vehicle-road was collected by dint of on-road experiments. The driving modes were divided into three patterns, namely, manual driving, warning assistant, and automatic driving, and the driver should self-report the selected driving modes. Meanwhile, the method of K-Means clustering was employed to cluster the driver's state into three levels based on the indexes of blood flow pulse (BVP) and skin conductance (SC). The current driving mode suitable to drivers was automatically clustered into three levels, depending on the fusing results of the drivers' self-reported and clustering. Then, the Ranker attributes sorting algorithm based on information gain was employed to select the key features, and three classification algorithms (SVM, NB, and KNN) were selected to confirm the irrelevant or redundant feature attributes. The results show that vehicle's speed, time headway, the distance of lane departure, the standard deviation of the front wheel angle, and drivers' experience are significantly influenced the selection of driving modes. The accuracy rates of support vector machine (SVM), Naive Bayes (NB) and K-nearest neighbor (KNN) algorithm are greater than 90%, when these five variables are selected as the feature subset. These five selected features are significantly different in different driving modes (p<0.05), except the value of vehicle speed and the distance of lane departure between warning assistant modes and automatic driving modes. The results can provide the theoretical support for the decision-making of driving modes of the intelligent vehicles.