汽车工程 ›› 2025, Vol. 47 ›› Issue (5): 809-819.doi: 10.19562/j.chinasae.qcgc.2025.05.002

• • 上一篇    下一篇

基于稀疏注意力的时空交互车辆轨迹预测

高凯1,2,刘欣宇2,胡林2(),黄向明1(),邹铁方2,刘鹏3   

  1. 1.湖南大学机械与运载工程学院,长沙 410082
    2.长沙理工大学汽车与机械工程学院,长沙 410114
    3.湖南星邦智能装备股份有限公司,长沙 410600
  • 收稿日期:2024-08-14 修回日期:2024-11-27 出版日期:2025-05-25 发布日期:2025-05-20
  • 通讯作者: 胡林,黄向明 E-mail:hulin@csust.edu.cn;h_xiangming@aliyun.com
  • 基金资助:
    国家杰出青年科学基金(52325211)、国家自然科学基金(52172399)和湖南省自然科学基金(2024JJ5023)资助。

Vehicle Trajectory Prediction with Spatial-Temporal Interaction Based on Sparse Attention

Kai Gao1,2,Xinyu Liu2,Lin Hu2(),Xiangming Huang1(),Tiefang Zou2,Peng Liu3   

  1. 1.College of Mechanical and Vehicle Engineering,Hunan University,Changsha 410082
    2.College of Automotive and Mechanical Engineering,Changsha University of Science & Technology,Changsha 410114
    3.Hunan Sinoboom Intelligent Equipment Co. ,Ltd. ,Changsha 410600
  • Received:2024-08-14 Revised:2024-11-27 Online:2025-05-25 Published:2025-05-20
  • Contact: Lin Hu,Xiangming Huang E-mail:hulin@csust.edu.cn;h_xiangming@aliyun.com

摘要:

在混合交通环境中,准确预测周边车辆轨迹对自动驾驶汽车安全至关重要。然而,现有技术在长时预测方面仍存在精度低和计算量大的问题。本文提出了一种结合意图概率的时空交互稀疏注意力模型,通过高效的编码-解码结构进行轨迹预测。模型首先构建位置掩码矩阵提取历史轨迹中的位置信息,利用稀疏注意力机制筛选出关键特征,并通过意图行为分析模块提高意图识别的准确率。最终将时空特征、位置特征和意图特征融合输入解码器,以多任务学习方式训练模型。试验结果表明,该模型在HighD和NGSIM数据集上相较于当前最优算法,在3~5 s长时预测的均方根误差均有降低,显著提升了预测效果。此外,通过实车试验对模型在实际场景中的表现进行验证,进一步展示了其在复杂交通环境中的应用潜力。

关键词: 交通工程, 轨迹预测, 稀疏注意力, 深度学习

Abstract:

In a mixed traffic ecosystem, accurately predicting the trajectories of surrounding vehicles is crucial for the safety of autonomous vehicles. However, existing technologies still face issues of accuracy and computational complexity in long-term prediction. A spatiotemporal interactive sparse attention model combined with intention probability is proposed in this paper, which predicts trajectories through an efficient encoder-decoder structure. The position mask matrix is first constructed to extract positional information from historical trajectories, and key features are selected using the sparse attention mechanism. The intention behavior analysis module is utilized to improve the accuracy of intention recognition. Finally, spatiotemporal features, positional features, and intention features are fused and input into the decoder, and the model is trained using a multi-task learning approach. The experimental results show that, compared to the optimal algorithm on the HighD and NGSIM datasets, the proposed model achieves a notable reduction in root mean square error (RMSE) in long-term prediction of 3 to 5 seconds, significantly enhancing prediction accuracy. In addition, the model's performance in real-world scenarios is validated through road tests, further demonstrating its application potential in complex traffic environment.

Key words: traffic engineering, trajectory prediction, sparse attention, deep learning