摘要
预测周围智能体的运动轨迹是实现自动驾驶行为决策规划的关键。面对复杂的车辆交互影响和多模态驾驶意图所带来的难题,本文提出一种基于车辆多目标交互行为建模的轨迹预测方法。该方法采用条件变分自编码器生成轨迹终点的多模态结果,结合自注意力机制和多头注意力机制来捕捉车辆之间的群体交互影响,最终使用逆强化学习输出多模态轨迹的最优决策,实现了同步预测多个目标轨迹。在高速公路数据集NGSIM上的实验结果证明该模型的有效性,并且预测效果整体优于现有方法。
Predicting trajectories of surrounding agents is critical to realize the decision-making planning of autonomous driving behaviors.Facing the difficulties brought by complex vehicle interaction and multimodal driving intention,this paper proposes a trajectory prediction method based on vehicle multi-agent interaction behavior modeling.The method uses conditional variational autoencoder to generate multi-modal results of the trajectory endpoints.By combination with the self-attention mechanism and multi-head attention mechanism,the influence of group interaction between vehicles is captured.Finally,the inverse reinforcement learning is used to output the optimal decision of multi-modal trajectory,realizing synchronous prediction of multi-agent trajectory.An experiment has been carried out on the NGSIM,which is a real-world trajectory prediction dataset on the highway traffic scene.The results prove effectiveness of the model,and the prediction effect is better than existing methods as a whole.
作者
赵靖文
李煊鹏
张为公
ZHAO Jingwen;LI Xuanpeng;ZHANG Weigong(School of Instrument Science and Engineering,Southeast University,Nanjing 210096,China)
出处
《智能系统学报》
CSCD
北大核心
2023年第3期480-488,共9页
CAAI Transactions on Intelligent Systems
基金
国家重点研发计划项目(2021YFB1600501)
国家自然科学基金项目(61906038)
中央高校基本科研业务费专项资金项目(2242021R41184).
关键词
轨迹预测
注意力机制
多目标交互
多模态预测
条件变分自编码器
端点生成
逆强化学习
决策校正
trajectory prediction
attention mechanism
multi-agent interaction
multimodal trajectory
conditional variational auto-encoder
endpoint generating
inverse reinforcement learning
decision refinement
