作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

基于图注意力机制的无地图场景轨迹预测方法

  • 发布日期:2023-11-28

A Graph Attention Mechanism-based Method for Trajectory Prediction in Map-Free Scenes

  • Published:2023-11-28

摘要: 随着5G、大数据和深度学习技术的发展,智慧交通领域的轨迹预测再次成为研究焦点。大量的真实轨迹数据集为准确的轨迹预测提供了数据基础。然而现有的轨迹预测工作大量依赖于高精地图,但高精地图的采集耗时长、成本高、处理复杂,难以快速适应智能交通的大面积普及。为解决无地图场景下车辆轨迹预测问题,提出了一种基于多模态数据时空特征的轨迹预测方法。构建了多个历史轨迹时空交互图,交叉使用时间和空间注意力并进行深度融合,以建模道路上车辆之间的时空关联性。最后,利用残差网络进行多目标多模态轨迹生成。在真实数据集Argoverse 2上进行了模型的训练和测试,实验结果表明,相较于其他先进的方法,该模型在单模态预测方面的各项指标提升了3.86%、3.89%、0.48%,在多模态预测方面的各项指标提升了0.78%、0.96%、0.42%。该方法能够有效地捕捉车辆移动轨迹的时间和空间特征,并可在自动驾驶等相关领域得到有效应用。

Abstract: With the development of 5G, big data, and deep learning technologies, trajectory prediction in the field of intelligent transportation has once again become a research focus. A large amount of real trajectory data sets provides a data foundation for accurate trajectory prediction. However, existing trajectory prediction work heavily relies on high definition map, which are time-consuming, costly, and complex to acquire, making it difficult to quickly adapt to the widespread adoption of intelligent transportation. To address the problem of vehicle trajectory prediction in map-free scenarios, a trajectory prediction method based on multi-modal data spatiotemporal features is proposed. Multiple historical trajectory spatiotemporal interaction graphs are constructed, and time and space attention are cross-utilized and deeply fused to model the spatiotemporal correlations between vehicles on the road. Finally, a residual network is used for multi-objective multi-modal trajectory generation. The model is trained and tested on the real dataset Argoverse 2, and the experimental results show that compared to other advanced methods, the model improves various indicators in single-modal prediction by 3.86%, 3.89%, and 0.48%, and in multi-modal prediction by 0.78%, 0.96%, and 0.42%. This method can effectively capture the temporal and spatial characteristics of vehicle movement trajectories and can be applied effectively in related fields such as autonomous driving.