汽车工程 ›› 2023, Vol. 45 ›› Issue (8): 1353-1361.doi: 10.19562/j.chinasae.qcgc.2023.08.006

所属专题: 智能网联汽车技术专题-规划&决策2023年

• • 上一篇    下一篇

基于车道目标引导的车辆轨迹预测

连静1,2,李硕贤2,刘一荻3,杨东方3,李琳辉1,2()   

  1. 1.大连理工大学,工业装备结构分析国家重点实验室,大连  116024
    2.大连理工大学汽车工程学院,大连  116024
    3.重庆长安汽车股份有限公司,重庆  400023
  • 收稿日期:2022-12-12 修回日期:2023-01-31 出版日期:2023-08-25 发布日期:2023-08-17
  • 通讯作者: 李琳辉 E-mail:lilinhui@dlut.edu.cn
  • 基金资助:
    国家自然科学基金(61976039);中央高校基本科研业务费专项资金(DUT22JC09);大连市科技创新基金(2021JJ12GX015)

Goal Supervised Attention Network for Vehicle Trajectory Prediction

Jing Lian1,2,Shuoxian Li2,Yidi Liu3,Dongfang Yang3,Linhui Li1,2()   

  1. 1.Dalian University of Technology,State Key Laboratory of Structural Analysis for Industrial Equipment,Dalian  116024
    2.School of Automotive Engineering,Dalian University of Technology,Dalian  116024
    3.Chongqing Chang’an Automobile Co. ,Ltd. ,Chongqing  400023
  • Received:2022-12-12 Revised:2023-01-31 Online:2023-08-25 Published:2023-08-17
  • Contact: Linhui Li E-mail:lilinhui@dlut.edu.cn

摘要:

有效融入车道线信息对准确预测车辆未来行驶轨迹具有重要意义。针对预测模型在融合车道线信息时存在效率低的问题,提出一种基于车道目标引导的车辆轨迹预测方法(GSA),在通过图网络融合车道段的几何及位置信息的基础上,从注意力模型出发,构建一种直接的车道目标点预测模块,由此监督模型将与车辆运动相关联的车道目标特征有效地融合到车辆的运动特征中,并考虑到周围车道拓扑结构随时间的变化。通过两个改进残差结构的Transformer网络依次提取低层运动特征以及融合车道目标点在时间尺度下的前后关联信息,逐步更新车辆运动特征。构建基于图网络的交互融合模块,使车辆运动特征在全局范围内聚合与传播。通过在Argoverse以及长安汽车轨迹预测数据集下的实验,验证了本文所提出的GSA方法能够有效提高复杂交通场景下车辆轨迹预测的精度和质量。

关键词: 轨迹预测, 地图特征编码, 目标引导

Abstract:

Effectively integrating lane information is significant for accurately predicting the future trajectory of vehicles. For the low efficiency problems existing in the fusion of lane information of the prediction model, a vehicle trajectory prediction method of Goal Supervised Attention (GSA) is proposed. Based on fusing the geometric and position information of the lane segment through the graph network, a lane goal prediction module is constructed in this paper starting from the attention model to directly supervise the model to fuse the lane goal features associated with vehicle motion into the vehicle’s motion characteristics while encoding changes in the surrounding lane topology over time. Through two Transformer networks with improved residual structure, low-level motion features are extracted and the correlation information of the lane goal at the time scale is fused sequentially to gradually update the vehicle motion features. An interaction fusion module based on a graph network is constructed to aggregate and propagate vehicle motion features globally. Experiments on the Argoverse and Changan vehicle trajectory prediction datasets show that the proposed GSA method can effectively improve the accuracy and quality of vehicle trajectory prediction in complex traffic scenarios.

Key words: trajectory prediction, map feature encoding, goal guided