不确定道路环境下基于人脑经验与理性的智能汽车行为决策方法研究

As one of the key technologies for the development of intelligent vehicle, maneuver decision making methods conforming to human driving style need to be studied in depth. The uncertain characteristics of the urban traffic environment caused huge challenges for maneuver decision making, such as the varied topological structure of the road, the flexibility motion of traffic participants and so on. While dealing with the familiar or unfamiliar driving scenarios, the human driver make a driving maneuver based on two different thinking processes, i.e. experiential decision making and rational reasoning. This mechanism can provide inspirations for the development of new decision making methods. Based on the analysis of human driver’s cognition and decision making processes, this project will propose a novel computing model of decision making by imitating the parallel cooperative experiential-rational mechanism of the human driver. At first, the high-dimensional discrete characteristics of the perception data are extracted to express the details of driving scenarios effectively. Then, the deep neural network is applied to achieve the nonlinear approximation of human drivers’ decision making ability by an end-to-end way, in which the raw perceived data are the inputs and the driving maneuvers are the outputs. Besides, the driving knowledge base will be established and the rational reasoning process will be simulated based on multi-level knowledge reasoning method. The conflict coordination mechanism will be established to choose the better decision making results based on aforementioned two processes. At last, the effectiveness of the proposed method will be verified through offline simulations and tested on public roadways. This project will not only promote the development of adaptable decision making methods, but also establish solid foundation for more advanced and acceptable intelligent vehicle.

研究符合人类驾驶习惯的行为决策方法是发展智能汽车技术的关键，城市道路环境存在结构多变、交通参与者机动性强等不确定因素，为行为决策方法的研究带来了挑战。人类在面向熟悉或陌生的驾驶场景时，表现出驾驶行为的经验决策和理性择优两种不同的思维过程，可以为研究新的决策方法提供借鉴。本项目将通过分析驾驶员的认知、决策过程，模拟人脑经验决策与理性推理双系统并行、协同决策的神经机理，建立行为决策计算模型；提取传感信息高维离散特征对驾驶场景细节进行有效表达，并设计原始传感信息端到车辆行为端的深度神经网络实现对人脑经验决策模型的非线性逼近；建立车辆行为知识库，基于多级知识推理方法实现对人脑理性推理过程的模拟，并建立冲突协调机制对两种决策结果进行择优；最后，采用离线仿真与实车实验的方式验证所提方法的有效性。本项目预期为提高智能汽车行为决策方法的适应性提供新途径，为车辆的高性能自主行驶建立技术基础。

研究符合人类驾驶习惯的行为决策方法是发展智能汽车技术的关键，城市道路环境存在结构多变、交通参与者机动性强等不确定因素，为行为决策方法的研究带来了挑战。人类在面向熟悉或陌生的驾驶场景时，表现出驾驶行为的经验决策和理性择优两种不同的思维过程，可以为研究新的决策方法提供借鉴。本项目从神经科学的角度对于人脑决策过程进行了研究，将人脑决策通路分为经验决策与理性推理系统，并明确了各决策通路的输入、输出及双系统输出的仲裁神经机制。同时，梳理了驾驶员的驾驶决策过程，其本质上是一种多源异构信息到驾驶行为的非线性映射。在此基础上，本项目模拟了人脑经验决策与理性推理双系统并行、协同决策的神经机理，建立了智能汽车行为决策计算模型。经验决策系统通过建立原始传感信息端到车辆行为端的深度神经网络实现了对人脑经验决策模型的非线性逼近，整个网络由特征融合子网络、特征提取子网络和行为预测子网络构成，实现了驾驶经验的有效提取，具备一定的泛化性能及记忆能力；同时，建立车辆行为知识库，定义了驾驶行为的效用函数，基于多级知识推理方法实现了对人脑理性推理过程的模拟，并使用优度评价法建立了冲突协调机制，对两种决策结果进行择优。本项目在《Journal of Advanced Transportation》、《2020 IEEE International Conference on Signal Processing, Communications and Computing》、《重庆交通大学学报》等国内外期刊会议上发表论文4篇（其中 SCI 检索 2篇），获软件著作权3项，正在申请发明专利1项，按时完成了既定目标。