基于驾驶人认知过程解析的城市交叉口穿越行为决策研究

The decision-making of crossing behavior at intersections, especially considering the strong time-space constraints and the multi-objective optimization, become one of the key issues which restrict the rapid development of the autonomous vehicles. The traditional rule-based decision models have a good real-time performance but poor generalization ability, while other decision models based on machine learning algorithms maintain a good generalization ability but a poor real-time performance. Thus, the decision-making model taking account of both generalization and real-time is of great significance for autonomous vehicles..Under complex and dynamic intersections, the driver has a good ability to cross. This study aims to analyze this cognitive process of human driver, and make decisions like human beings when autonomous vehicle crossing intersections. Firstly, by mining the driving trajectory data and analyzing individual attention distribution pattern, which can contribute to understand the relationship between the driver’s attention and traffic environmental features related to driving tasks, and the interest area of decision, which means the key environmental features that have the greatest impact on decision-making in human driver’s cognitive process, can be obtained and learnt. Therefore, these Interest Area of Decision can make a guarantee for real-time decision of autonomous vehicles by reducing the dimensions of the state space. This paper discusses the internal relationship between environmental differences and driving expectations based on cognitive psychology so as to help autonomous vehicles make decisions in line with human expectations in current states and improve autonomous vehicles’ cognitive ability. Furthermore, in order to improve the generalization ability and real-time performance of the decision algorithm, the multi-objective optimization decision model, which takes safety, efficiency and comfort into consideration, is built in combination with rule-based decision models and reinforcement learning method. This model can use on-line decision network to make the desired decision based on off-line decision results from rule-based models and reinforcement learning network. The off-line results can help on-line network improve the convergence rate and reduce the randomness of decisions. The study findings can provide the decision basis for autonomous vehicles under the complex traffic environment. Moreover, it can also provide theoretical support for the research of autonomous vehicles based on driver cognition theory.

强时空约束、多目标寻优的穿越行为决策是制约智能车辆快速发展的关键问题之一。传统规则模型实时性好、泛化能力差；机器学习算法有一定泛化能力、但实时性差。研究兼顾泛化能力和实时性的决策模型对于智能车辆实现高效行驶具有重要意义。.复杂城市环境下，驾驶人具备良好的穿越行为能力，本研究试图基于驾驶人认知过程解析，建立穿越行为“仿人”决策模型。首先，通过群体行车轨迹数据挖掘和个体注意力分配模式分析，获取行为决策兴趣点，实时更新决策状态空间，对状态空间降维，为决策实时性提供保证。同时，基于认知心理学剖析环境差异对驾驶期望的影响机理，提升智能车辆认知能力。最后，融合模型驱动和强化学习两种算法优势，提供离线决策的待选结果，指导在线决策网络完成满足通行期望的仿人决策，降低决策随机性，保证算法泛化能力和实时性。研究成果不仅可以为智能车辆“仿人”通行提供理论依据，也可以为认知心理学在智能车辆的深度应用提供理论支撑。

强时空约束、多目标寻优的穿越行为决策是制约智能车辆快速发展的关键问题之一。传统规则模型实时性好、但泛化能力差；机器学习算法有一定泛化能力、但实时性差。研究兼顾泛化能力和实时性的决策模型对于智能车辆实现高效行驶具有重要意义。驾驶人在复杂城市环境下具备良好的穿越行为能力，本研究基于驾驶人认知过程解析，建立穿越行为“仿人”决策模型。首先，通过群体行车轨迹数据挖掘和个体注意力分配模式分析，获取行为决策兴趣点，实时更新决策状态空间，对状态空间降维，为决策实时性提供保证。同时，基于认知心理学剖析环境差异对驾驶期望的影响机理，提升智能车辆认知能力。最后，融合模型驱动和强化学习两种算法优势，提供离线决策的待选结果，指导在线决策网络完成满足通行期望的仿人决策，降低决策随机性，保证算法泛化能力和实时性。.具体工作包括：提出了基于群体行车轨迹和深度学习网络的决策兴趣区域获取方法；界定了有人与无人驾驶车辆的路段和交叉口交互行为；提出了基于周边多车交互和驾驶认知模式分析的轨迹预测方法；建立了混行条件车辆协同换道模型；提出了基于冲突消解的交叉口左转通行决策模型和方法；提出了交叉口LTAP-OD场景交通状态估计和POMDP决策方法；提出了基于元探索DDPG的交叉口无人车多目标优化决策方法。经过算法验证表明：1）行车意图预测综合准确率达92%，5s内的横向轨迹预测误差在0.3m内；2）对车辆换道及交叉口穿越过程中合作竞争程度预测准确率超过87.3%和91.6%。研究成果不仅可以为智能车辆“仿人”通行提供理论依据，也可以为认知心理学在智能车辆的深度应用提供理论支撑。.本课题提供复杂动态城市交叉口仿人穿越行为决策模型2套，并进行了虚拟仿真和实际场景测试应用，发表SCI论文6篇，EI论文4篇，申请发明专利8项，专著1项，参加国内外学术研究总结与进展报告4项，培养工程博士研究生2名，硕士研究生7名。超额完成各项指标。