智能车个性化学习方法与行为决策模型研究

Although the development of intelligent vehicle is progressing rapidly, automatic driving and manual driving living together to control vehicles will be an inevitable stage for a long time. The existing autonomous driving system of intelligent vehicles is generally based on fixed rules of design, which is difficult to meet drivers’ needs since each person has different driving habits. Intelligent driving will be improved if intelligent vehicles can independently learn the driving habits of the current driver and develop individualized driving decision-making under various constraints such as traffic rules and collision avoidance rules. To tackle the above-mentioned problems, this study is mainly based on the theory of artificial intelligence. Firstly, the transfer learning theory will be used to develop independent learning method of the decision-making of driving behavior to achieve the goal of learning the current driver’s driving characteristics. Then, based on the theory of information entropy, the driving behavior decision law of a typical driving style is used to provide the restraint for decision-making strategies of personalized driving behavior. Finally, a decision-making model of intelligent vehicles based on sparse representation under multi-constraint condition is constructed, and some optimization methods will be used to solve the model. The results of this study are useful for providing new ideas and methodologies for both advanced driver assistant systems and autonomous driving system.

智能车技术虽然发展迅速，但将在很长一段时期内处于人机共驾阶段。现有的智能车自主驾驶系统都是同一模型，没有考虑不同风格驾乘人员的感受。如果智能车系统能够自主学习当前驾驶者的驾驶行为习惯，并在交通规则、避碰规则等多种约束下采取个性化的行为决策，将会提高智能车的自主驾驶技术水平。课题基于人工智能理论开展研究，首先，利用迁移学习理论研究驾驶行为决策自主学习方法，实现智能车系统自主学习当前驾驶者的驾驶特性。然后，基于信息熵理论挖掘典型驾驶风格的驾驶行为决策规律，为智能车行为决策提供个性化的约束。最后，构建多约束条件下基于稀疏表示的智能车个性化行为决策模型，并利用最优化理论设计模型求解方法获取个性化行为决策。研究成果可为自动驾驶技术和辅助驾驶技术提供新思路和新方法。

智能车技术虽然发展迅速，但智能车将在很长一段时期内处于人机共驾阶段。现有的智能车自主驾驶系统都是同一模型，没有考虑不同风格驾乘人员的感受。因此，本课题紧紧围绕智能车人性化驾驶理念，以实现智能车个性化驾驶决策为目标，基于改进型马尔科夫毯的特征选择算法，建立了驾驶特征提取方法，选用车速、车辆横向偏移距离和角度以及车辆位置信息作为驾驶特征选择结果。深入研究了驾驶人驾驶个性的表征方法，提出了表征驾驶个性的驾驶指纹概念的模型及其构建方法，采用驾驶指纹图描述驾驶个性。基于隐性狄利克雷算法构建了个性化驾驶风格识别模型，实现了对驾驶人驾驶风格的理解。综合运用车辆动力学、人工势场、强化学习等方法，综合考虑行驶特征、交通规则、避碰规则、个性化驾驶行为决策等约束，构建了智能车多约束个性化行为决策模型，基于最优化理论设计模型求解方法，在满足驾驶人驾驶风格条件下，模型具有良好的横向控制效果。课题从理论和实践上为自动驾驶技术和辅助驾驶技术提供新思路和新方法，能够促进智能车的应用和推广。