基于结构化道路的车辆多模态轨迹预测方法  

Multi-modal Vehicle Trajectory Prediction Method Based on Structured Road

作　　者：胡杰[1] 吴作伟 张志凌 赵文龙 代怡鹏 HU Jie;WU Zuo-wei;ZHANG Zhi-ling;ZHAO Wen-long;DAI Yi-peng(School of Automotive Engineering,Wuhan University of Technology,Wuhan 430070,Hubei,China;Commercial Product R&D Institute,Dongfeng Automobile Co.Ltd.,Wuhan 430100,Hubei,China)

机构地区：[1]武汉理工大学汽车工程学院,湖北武汉430070 [2]东风汽车股份有限公司商品研发院,湖北武汉430100

出　　处：《中国公路学报》2025年第2期286-295,共10页China Journal of Highway and Transport

基　　金：2023年湖北省重大攻关项目(JD)(2023BAA017)。

摘　　要：车辆轨迹预测是自动驾驶系统的核心功能之一,是下游决策规划模块做出安全有效的驾驶行为的重要基础。为实现结构化道路场景下自动驾驶汽车对周围车辆长时域准确轨迹预测,在轨迹预测经典模型VectorNet的基础上,提出了一种分层交互的车辆多模态轨迹预测方法S-VectorNet。首先,引入门控循环单元(Gated Recurrent Unit,GRU)编码历史轨迹信息和地图信息,提升了编码特征的时间表征能力;其次,构建了一种基于注意力块和图神经网络(Graph Neural Networks,GNN)的双层交互模型对交通主体(包括目标车辆和周围交通主体)与地图间、交通主体相互间的交互作用建模,实现了更好的长程动态交互建模能力;然后,提出了一种随时间动态更新的场景表示模块,通过多头注意力机制和时间序列模型捕捉个体运动状态和交互的时间相关性,使模型学习到丰富的场景记忆信息;最后,在多模态轨迹生成方面使用两阶段轨迹生成方法,提高了模型对预测端点的捕捉能力。在公开数据集Argoverse上进行的试验表明:S-VectorNet在验证集上较基准模型最小平均位移误差降低12%,最小最终位移误差降低22%;在测试集上最小平均位移误差为0.83 m,最小最终位移误差为1.23 m,与现有其他轨迹预测模型相比综合性能优势明显。Vehicle trajectory prediction is a core function of autonomous driving systems and serves as a critical foundation for downstream decision-making and planning modules,enabling safe and effective driving behaviors.To achieve accurate long-term trajectory prediction of surrounding vehicles in structured road scenarios,a hierarchical,interactive vehicle multi-modal trajectory prediction method,S-VectorNet,was proposed based on the classical VectorNet model.First,gated recurrent unit(GRU)was introduced to encode historical trajectory data and map information,thereby enhancing the temporal representation capability of the encoded features.Second,a two-layer interaction model incorporating attention blocks and graph neural network(GNN)was constructed to model interactions between traffic agents(target vehicles and surrounding agents)and the map.This approach improves the model's ability to capture long-range dynamic interactions.Next,dynamic scene representation module,which is updated over time,was proposed to capture the temporal correlations of individual motion states and interactions using multi-head attention mechanisms and time-series models.This allows the model to learn rich scene memory information.Finally,a two-stage trajectory generation method was employed to generate multi-modal trajectories,enhancing the model's ability to predict trajectory endpoint.Experiments conducted on the Argoverse dataset show that S-VectorNet reduces the minimum average displacement error by 12%and the minimum final displacement error by 22%compared to the baseline model on the validation set.On the test set,the minimum average displacement error is 0.83 m,and the minimum final displacement error is 1.23 m,demonstrating significant comprehensive performance advantages over other existing trajectory prediction models.

关 键 词：汽车工程 轨迹预测 深度学习 自动驾驶汽车 交互建模 

分 类 号：U463.6[机械工程—车辆工程]