基于图时空注意力的多车交互轨迹预测模型  

Multi-Vehicle Interaction Trajectory Prediction Model Based on Graph Spatial-Temporal Attention

作　　者：张新锋[1,2] 赵娟 刘国华 刘鹏菲 Zhang Xinfeng;Zhao Juan;Liu Guohua;Liu Pengfei(School of Automobile,Chang’an University,Xi’an 710064;School of Transportation and Logistics Engineering,Xinjiang Agricultural University,Urumqi 830052)

机构地区：[1]长安大学汽车学院,西安710064 [2]新疆农业大学交通与物流工程学院,乌鲁木齐830052

出　　处：《汽车技术》2025年第3期30-38,共9页Automobile Technology

基　　金：陕西省重点研发计划项目资助(2022GY-303);西安市科技计划项目资助(2022GXFW0152)。

摘　　要：为有效提取高速交通场景下车辆间的交互特征,从而准确预测动态障碍轨迹,基于编-解码框架,提出基于图时空注意力的多车交互轨迹预测模型。结合斥力场和图模型建立车-车图交互场,利用节点和邻接特征矩阵表征车辆之间的动态交互,通过图空间注意力和时间多头注意力提取深层时空交互,获取图时空融合编码;将车辆横纵向行为意图独热编码与其拼接,实现目标车辆多模态轨迹预测。利用NGSIM数据集进行验证,相较于其他6种模型,该模型RMSE和NLL值最低;通过消融实验进一步验证图交互场的有效性,结果表明,该模型能够有效提高车辆轨迹预测精度。In order to effectively extract interaction features among vehicles in high-speed traffic scenarios,thus accurately predict the trajectories of dynamic obstacles,this paper proposes a multi-vehicle interaction trajectory prediction model using the coding-decoding framework based on the graph spatial-temporal attention mechanism.The vehicle-to-vehicle graph interaction field is established by combining the repulsive force field and the graph model,the node feature matrix and the adjacency feature matrix are used to characterize the dynamic interaction between the vehicle and the surrounding vehicles,and the deep spatial-temporal interaction features are extracted by the graph spatial attention and temporal polytope attention to obtain the graph spatial-temporal fusion coding features.The one-hot encoding of the longitudinal and lateral behavior intentions of the vehicles is concatenated with the encoding to achieve multimodal trajectory prediction for the target vehicles.Validation using the NGSIM dataset shows that,compared with 6 other models,the proposed model achieves the lowest RMSE and NLL values.Ablation experiments further validate the effectiveness of the graph interaction field,demonstrating that the model can significantly improve the accuracy of vehicle trajectory prediction.

关 键 词：多车交互 斥力场 注意力机制 图模型 轨迹预测 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]