行人-车辆异构多智能体的群集避撞行为机理研究

Facing to the conflict relationship between pedestrians and intelligent vehicles in complex mixed traffic environment, this project intends to take the multi-agent system theory as the main line and aims to study the relationship between external characteristics and internal mechanism of behavior, the relationship between microscopic individual interactive and macroscopic collective phenomena by: proposing a target detection and behavior state analysis method combining deep-inductive learning and high-degree cubature kalman filter algorithm, defining the impact mechanism between collision avoidance behavior and behavioral state; proposing the decision making model of cluster collision avoidance behavior based on multi-agent reinforcement learning, analyzing the influence of behavioral state, behavioral psychology, communication delay, traffic environment and other factors to the decision-making process of cluster collision avoidance behavior, revealing the formation mechanism of cluster collision avoidance behavior; proposing the pedestrian-vehicle cluster dynamics model, studying the emergence process of micro-individual behavior to macro-cluster phenomenon, revealing the evolutionary law of cluster collision avoidance behavior; rebuilding the pedestrian-vehicle heterogeneous multi-agent cluster motion simulation model system, proposing a safety evaluation method for cluster collision avoidance behavior and developing empirical research. This project provides theoretical basis for intelligent vehicle active safety system optimization, pedestrian illegal traffic behavior correction, and road infrastructure safety improvement.

面向复杂混合交通环境中行人与智能车辆的冲突关系，本项目以多智能体系统理论为主线，从行人-车辆异构多智能体的外在行为特征到内在行为机理、微观个体作用到宏观群集现象的关系入手：提出深度-迁移学习与高阶容积卡尔曼滤波相结合的目标检测与行为状态解析方法，明确群集避撞行为本身与行为状态的影响机制；提出基于多智能体强化学习的群集避撞行为决策模型，分析行为状态、行为心理、通信时延、交通环境等因素对群集避撞行为产生过程的影响，揭示群集避撞行为的形成机理；提出行人-车辆群集动力学模型，探索微观个体通过交互作用到宏观群集现象的涌现过程，揭示群集避撞行为的动力学演化规律；重构行人-车辆异构多智能体的群集运动仿真体系，提出群集避撞行为的安全性评价方法，并开展实证研究。为智能车辆主动安全系统优化、行人违法交通行为矫正、道路基础设施安全性改善等提供理论依据。

面向我国复杂混合交通环境中行人与智能车辆的碰撞避免问题，本项目深入探索行人-车辆异构多智能体的群集避撞行为机理，并对其中关键技术进行深入研究。从行人-车辆目标检测与识别、行人-车辆目标跟踪与容错控制、行人-车辆目标轨迹预测、行人-车辆目标行为冲突下的通行策略优化等几方面入手，提出了基于多模态特征融合的低辨识目标检测方法、基于图像和点云实例匹配的目标状态估计和跟踪、基于容错自适应模型预测的自主车辆路径跟踪控制、基于卡尔曼滤波与改进社会力模型的行人轨迹预测、基于演化博弈论的行人-车辆目标通行策略优化等先进算法，并且构建了复杂KITTI数据集、行人轨迹预测数据集等，用于验证所提出方法的有效性和合理性。此外，还作为核心成员申报了江苏省智能电动运载装备工程研究中心，为本研究的顺利开展提供了必要的软硬件支持。产出一系列科研成果，培养了多名研究生和博士顺利毕业。本项目的研究内容为智能驾驶和智慧交通领域最前沿的研究方向，项目的成功执行将为智能车辆行为产生系统优化、行人违法交通行为矫正、道路基础设施安全性改善等提供理论依据。