仿驾驶员态势感知模型与人机协同驾驶决策研究

With increasing of vehicles in China, road traffic accidents happen rather frequently, and statistics show that almost 75% traffic accidents are caused by the driver distractions, driving fatigues or misjudgments. Currently, the driver behavior factors are rarely considered in most of the automobile active safety systems. Thus, how to evaluate the driver behaviors properly is crucial, but difficult for the vehicle active safety system. Some studies show that the driver’s manipulations, facial changes (eyes, lips) and his/her psychological and physiological reactions vary with the traffic scenarios, and there exists certain regularities to follow. Therefore, on basis of the present studies about driver’s manipulation intention analysis and driver visual attention analysis, a multi-dimensional assessment of the driver’s behavioral features, including the driver’s visual attention, mental overloads, manipulation intentions and driving skills, is explored and proposed by deep learning method in this project. When the drivers are facing with different traffic scenes, the relationships between the driver behaviors and the traffic scene are further studied and analyzed. Then a mimic-driver situation awareness model is built accordingly. Meanwhile, a decision-making strategy of human-vehicle collaborative drive and its evaluation are established, combining with the features and advantages of human intelligence (HI), artificial intelligence (AI) and computer technology (CT). It provides a theoretical basis and key technology for automobile intelligent driving, which includes assistance driving and automatic driving.

随着我国汽车保有量持续增加，道路交通事故频发，其中由于驾驶员的注意力分散、疲劳驾驶或错误判断等导致的交通事故约占75%。目前大多数主动安全系统很少考虑驾驶员行为因素，因而如何评估驾驶员行为是制约此类安全系统的核心和难点。研究表明面对不同的交通场景时驾驶员的操作行为、面部（即眼睛和嘴唇）变化以及心理生理反应等都随之变化，且有一定规律可循。因此本项目将立足于驾驶员操作意图分析、驾驶员视觉注意力分析等已开展的研究基础，采用深度学习方法，探索驾驶员在不同危险程度的交通场景中，其视觉注意力、精神负荷量、操作意图和驾驶技能等驾驶员行为多维度评估机制，以及驾驶员行为特征与交通事故的关联关系，并以此建立仿驾驶员态势感知模型。同时，充分发挥人的智能HI、人工智能AI和计算机技术CT各自的优势及特点，研究人机协同驾驶决策机制与评估体系，为汽车的智能驾驶（辅助驾驶和自动驾驶）储备理论基础和关键技术。

本项目基于驾驶员在环仿真平台，综合考虑了交通环境信息、车辆状态信息、驾驶员生理信号、心理信号、眼动信号等人-车-路多维信号，采集了驾驶员行为原始数据信息，建立了驾驶员行为数据库，使用支持向量机、半监督学习方法等机器学习方法建立了驾驶行为多维评估模型，并探究了驾驶员执行次要驾驶任务时，驾驶员人口学因素与碰撞危险之间的关联关系。此外，本项目进一步研究了人机协同控制决策机制，基于驾驶换道意图和环境危险评估等信息提出了多种人机协同控制策略，解决了驾驶员与自动驾驶系统之间的控制权冲突问题，实现了人机协同避障，取得良好的车辆稳定性能，为智能驾驶系统设计提供了合理思路和理论方法。