汽车工程 ›› 2024, Vol. 46 ›› Issue (3): 396-406.doi: 10.19562/j.chinasae.qcgc.2024.03.003

• • 上一篇    

城市道路场景下考虑多类交通参与者的多模态车辆轨迹预测方法研究

周亦威1,2,夏莫1,朱冰3()   

  1. 1.上海理工大学管理学院,上海 200093
    2.上海理工大学智慧应急管理学院,上海 200093
    3.吉林大学,汽车仿真与控制国家重点实验室,长春 130022
  • 收稿日期:2023-08-13 修回日期:2023-11-22 出版日期:2024-03-25 发布日期:2024-03-18
  • 通讯作者: 朱冰 E-mail:zhubing@jlu.edu.cn
  • 基金资助:
    教育部人文社会科学青年基金项目(22YJC790189);上海纽约大学上海市城市设计与城市科学重点实验室开放课题基金(Grant 2023 YWZhou_LOUD)

Multimodal Vehicle Trajectory Prediction Methods Considering Multiple Traffic Participants in Urban Road Scenarios

Yiwei Zhou1,2,Mo Xia1,Bing Zhu3()   

  1. 1.Business School,University of Shanghai for Science and Technology,Shanghai  200093
    2.School of Intelligent Emergency Management,University of Shanghai for Science and Technology,Shanghai  200093
    3.Jilin University,State Key Laboratory of Automotive Simulation and Control,Changchun  130022
  • Received:2023-08-13 Revised:2023-11-22 Online:2024-03-25 Published:2024-03-18
  • Contact: Bing Zhu E-mail:zhubing@jlu.edu.cn

摘要:

车辆轨迹预测是自动驾驶的关键技术之一,针对以往模型较少考虑城市道路场景中车辆以外多类交通参与者的问题,本研究提出了一种多类交通参与者的多模态车辆轨迹预测模型。该模型使用门控循环单元对历史轨迹信息进行编码,并利用注意力机制将多类交通参与者的特征映射到用图结构表达的驾驶场景中,通过图注意力网络进行环境特征提取,从而使模型能感知环境中的多类交通参与者。此外,模型通过节点轨迹预测与坐标轨迹预测模块输出最终的多模态轨迹预测结果。基于城市道路场景数据集nuScenes的实验结果表明:相较于同类现有模型,所提出的模型算力需求更低、预测更准确,且能适用于人车混合的城市道路驾驶场景。

关键词: 车辆轨迹预测, 自动驾驶, 图注意力网络, 注意力机制, 门控循环单元

Abstract:

Vehicle trajectory prediction is one of the key technologies for autonomous driving. In view of the problem that previous prediction models rarely consider multiple types of participants other than vehicles in urban road scenarios, a multimodal vehicle trajectory prediction model considering multiple types of traffic participants is proposed in this paper,. The historical trajectory information is encoded by gated recurrent units, and the features of multiple types of traffic participants are mapped into the driving scene expressed by a graph structure through the attention mechanism, with the context feature extracted by the graph attention network, so that the model can perceive different traffic participants in the environment. In addition, the final multimodal trajectory prediction results are output through the node trajectory prediction module and the coordinate trajectory prediction module. Experiments on nuScenes, a dataset under urban road scenarios, show that, compared to similar existing models, the model proposed has lower computational requirements and more accurate prediction, which is applicable to urban road driving scenarios with mixed human and vehicle traffic.

Key words: vehicle trajectory prediction, autonomous driving, graph attention network(GAT), attention mechanism, gated recurrent unit(GRU)