基于驾驶人行为和车辆运行状态变化的驾驶分心识别方法研究

Driving distraction (such as the use of in-vehicle information systems or mobile phone) will endanger traffic safety seriously. A recent foreign study has shown that nearly 80 percent of all crashes and 65 percent of all near-crashes contain at least one type of driving distraction. In this work, we intend to carry out real car driving tests both in the automotive testing ground and on the road, and acquire drivers' eye movement and operating behavior data, together with vehicle motion state data in the driving processes with and without distraction. The changes in the characteristics and difference of these parameters will be analyzed in-depth when the primary driving task is impacted by secondary driving tasks with visual distraction, operating distraction, cognitive distraction and/or their combinations. By using significant test and the transformation based on the multiwavelet theory of test parameters, the impact of distraction-irrelevant factors and will be peeled, and the driving distraction characteristic parameters set be extracted. Finally by using the kernel principal component analysis (KPCA) method and artificial neural network (ANN) technology, the sensitivity of each characteristic parameter for different driving distraction form and its combinations will be analyzed. Based on sparse Bayesian theory and the four-element evaluation method of DHGF, the identification results of distraction identification sub-module will be analyzed synthetically, and the driving distraction identification model be established. The driving distraction identification method can provide theoretical support for the development of an early-warning system for distraction.

驾驶分心（如开车时使用车内信息系统或手机）会严重威胁交通安全，国外最新研究表明，近80%的碰撞和65%的临界碰撞都与驾驶分心有关。本项目拟在汽车试验场和实际道路上开展实车驾驶试验，实时采集驾驶人在分心和未分心驾驶过程的眼动数据、操作行为数据与车辆运动状态数据，深入分析这些参数在受到视觉分心、操作分心、认知分心等驾驶分心次任务及其组合影响时的变化特点与差异；通过对测试数据进行显著性检验，并基于多小波理论对测试参数进行变换，剥离与分心因素不相关的其他信息的影响，从而提取驾驶分心特征参数集；最后采用核主成分分析法和人工神经网络技术分析各特征参数对于不同驾驶分心形式及其组合的敏感度，基于稀疏Bayesian理论并运用DHGF四元评价法对各分心识别子模块的识别结果进行综合处理，建立驾驶分心识别模型。所提出的驾驶分心识别方法能为分心预警系统的开发提供理论支撑。

基于多资源理论、事故倾向性理论和驾驶适宜性检测，搭建了驾驶次任务等级测试平台。通过测量22位被试者的反应时间、操作错误率和任务完成时间，采用聚类分析法对驾驶分心等级和危险程度进行合理匹配，划分了驾驶次任务分心等级。结果发现：参照第0级未分心的情况，所有驾驶分心次任务对驾驶绩效的影响可划分为1~5等级，分别为轻度分心、中轻度分心、中度分心、重度分心和超重度分心。. 采用时域分析和数理统计等方法分析不同认知分心等级下驾驶人视觉行为、操作行为和车辆运动状态的变化特征，并提出认知分心敏感特征参数集。结果表明，眨眼频次、注视道路中央区域的百分比值以及视角标准差这3个眼动参数对于驾驶人认知分心非常敏感；转向盘转向熵、转向盘角速度均值、转向盘角速度标准差、转向盘持续静止时长和油门踏板持续静止时长等操作行为参数对认知分心敏感；车速均值、车速标准差、横向偏移距离均值、横向偏移距离标准差和横向偏移距离幅度等车辆运动参数也对认知分心敏感。. 利用小波包分析方法对所提取的认知分心敏感参数进行小波包分解重构，研究各参数在不同驾驶状态下各频带相对能量的分布特点，提取能够表征认知分心的特征频带。经降维约减得到认知分心综合表征指标，在此基础上运用支持向量机实现对认知分心的离线检测。结果表明，各参数主要集中在低频区域，认知分心时有向较高频区域移动的趋势，且认知分心时各参数幅值均有显著增加。基于降维得到的5个综合指标所构建的支持向量机模型，可以较好地实现对认知分心行为的离线检测，其综合识别率达到84.97%。另外，本研究还在驾驶模拟器上开展了认知分心实验，运用随机森林方法构建了认知分心识别模型，应用网格搜索确定最优模型参数，并采用100次留出法评估模型性能。结果表明:当决策树数量为125、最大特征数为5时，模型识别平均准确率为83. 69%。