为实现安全、高效的智能决策，智能汽车需要理解其他道路使用者的行为意图及其演变规律，准确评估交通环境中的行车风险。然而，在机非混行的交通环境中，现有研究难以对动态、随机的弱势道路使用者进行准确的行为建模，难以全面地评估人、车和环境耦合的行车风险，难以实现复杂场景下的实时决策。针对上述问题，本文开展了基于行人行为预测和人-车-环境耦合行车风险评估的智能汽车换道决策方法研究。首先，提出了融合概率图模型与深度学习模型的行人行为预测方法。基于动态贝叶斯网络模型通过概率推理预测行人的意图以及状态分布，通过深度学习模型拟合行人的历史轨迹与预测轨迹的回归关系。提出在线自适应的模型权重调节算法对动态贝叶斯网络模型和深度学习模型进行动态加权，实现复杂场景下概率图与深度学习技术的优势互补，提高了行人动态随机行为的预测精度。其次，建立了基于场论思想的人-车-环境耦合行车风险评估模型。根据人、车和环境三要素与行车风险的因果关系建立行车风险致因模型，考虑了碰撞、经验和违规行为主导的行车风险；提出“行车风险场”基本概念并建立行车风险量化模型；计算行车风险阈值，基于目标行为预测估计碰撞概率，提出相对行车风险指数，实现了人、车、环境耦合行车风险的量化、统一和预测。然后，提出了基于最优化模型和模仿学习的分层换道决策方法。首先，基于行为预测和行车风险场模型约束建立换道时机优化模型；其次，基于行车风险逻辑约束建立混合整数二次规划模型生成换道参考路径；然后，利用最优控制对参考路径进行跟踪，采集路径对应的状态与最优控制量建立样本数据库，通过改进的数据聚合算法在线更新数据库并训练深度神经网络模型，通过神经网络在线调用生成驾驶控制量，从而提高了自动驾驶换道决策的实时性。最后，开展实车与仿真试验。基于实车数据和仿真验证了行人行为预测方法、行车风险场模型和分层换道决策方法的有效性。

To enable safe and efficient decision-making for intelligent vehicles, it is essential to assess how critical the future situation might be by predicting the most likely evolution of the current traffic situation. However, it is a non-trivial task to accurately predict the dynamic and stochastic behavior of vulnerable road users (VRUs) for intelligent vehicles on urban roads. Also, the driving risks formed by the coupling interactions among driver, vehicle, and environment are difficult to be quantified and evaluated with a unified framework. Besides, the dynamic changes in traffic scenes make it challenging for intelligent vehicles to generate effective driving decisions that guarantee safety, efficiency, and driving comfort simultaneously. In this study, a hybrid framework that combines probabilistic graphical model and sequence learning is proposed for pedestrian motion prediction. Based on that, a potential field-based risk assessment model that considers interactions among driver-vehicle-environment is established. Then, an imitation learning method that integrates the mixed-integer quadratic planning and supervised learning is proposed for the decision-making of automated lane change. First, this study presents a hybrid framework for pedestrian trajectory prediction, which integrates Dynamic Bayesian Networks (DBN) and Sequence-to-Sequence (Seq2seq) model through an adaptive online weighting method, i.e., DBN-Seq2Seq. DBN utilizes environmental features and kinematic information to infer the pedestrian’s motion intentions through probabilistic reasoning. Seq2seq views trajectory predictions as sequence generation tasks, in which the future trajectories are generated relying on the observed trajectories. A real-world pedestrian motion dataset is employed for model validations and it is enlarged through data augmentation techniques to enable training of data-driven approaches. We compare our model with several typical baselines methods and results show that our model outperforms those baselines. This method expects to provide a basis for intention recognition and motion prediction of vulnerable road users at un-signalized intersections. Then, a driving risk assessment method, i.e., the driving risk field model is proposed based on the potential field theory, which quantifies the driving risks caused by the complicated coupling and interactions among driver, vehicle and environment. Based on the analyses of the driver’s risk perception mechanism and driving risk causalities, the driving risk field model considers the risks that are related to the potential collisions, the abnormal driver behaviors and the driver’s driving experience. A driving risk index is proposed for risk quantification in the driving risk field model and makes it possible to round up the driving risks of different traffic elements to the same scale. With the intention recognition and motion predictions of traffic participants, the collision probability is calculated and used for modifying the driving risk index, in order to generate a proactive and predictive risk assessment. This method expected to address the challenges of risk quantification and unified assessment for different traffic elements in dynamic traffic scenes. A hierarchical decision-making method for automated lane change of intelligent vehicles on the highway is proposed based on imitation learning, which combines the optimization theory and supervised learning. Considering the driving risks affected by the lane marks, the constraint logic programming is employed for the path planning of lane change, which models the driving risks caused by vehicles on multiple lanes as the logic constraints. Also, the driving risk constraints are represented with the proposed driving risk field model. Then, the logic programming problem of lane change path planning is established as the MIQP (mixed-integer quadratic programming) model. With the reference path provided by MIQP, an MPC (model predictive controller) is used for path tracking and then the state inputs and control outputs from MPC are recorded as the data samples for supervised learning. The supervised learning is actually employed to learn the mapping from the observations to the actions based on training on large amounts of data. A DNN (Deep Neural Network) is established based on Keras for the supervised learning and then is trained offline with the collected data samples. To enable robust imitation learning, a DAgger algorithm is employed to retrain the DNN model online. This method expects to provide a safe, efficient and scalable decision-making strategy for intelligent vehicles. Finally, real-world driving experiments and simulations are conducted for model verifications. With the real-world pedestrian motion dataset, the proposed DBN-Seq2Seq model can achieve state-of-the-art performances for pedestrian trajectory prediction task. Also, typical traffic scenarios are extracted from the naturalistic driving dataset for validation, and the proposed driving risk field model can generate accurate, predictive and proactive risk assessments, which is more consistent with the risk perception of human drivers. Moreover, the imitation learning-based motion planning model is validated by numerical simulations and real-world lane change scenarios. Results show that the proposed method gains more efficiency than conventional optimization-based methods, and has stronger scalability in dynamic and complicated scenarios