基于仿射全注意力机制的车辆轨迹预测方法

A Vehicle Trajectory Prediction Method Based on Affine Full Attention Mechanism

摘要: 针对现有车辆轨迹预测方法在复杂交通场景中时空信息提取不充分、多智能体交互建模能力有限的问题，本文提出一种基于仿射全注意力机制的车辆轨迹预测方法BiFSTNet。该方法采用动态邻接矩阵建模车辆间空间关系，利用视觉Transformer从多车辆历史轨迹中提取全局上下文特征，并设计仿射全注意力机制融合时间多头注意力和空间图注意力网络，有效捕获时空交互模式和长期时间依赖关系。采用双向长短期记忆网络构建编码器–解码器架构，实现端到端的轨迹预测。在NGSIM US-101和I-80数据集上的实验结果表明，BiFSTNet在US-101数据集上平均均方根误差为1.77m，在I-80数据集上平均均方根误差为2.06m，相比基线方法在各预测时间范围内均有性能提升。

Abstract: To address the insufficient spatiotemporal information extraction and limited multi-agent interaction modeling capability of existing vehicle trajectory prediction methods in complex traffic scenarios, this paper proposes BiFSTNet, a vehicle trajectory prediction method based on affine full attention mechanism. The proposed method employs dynamic adjacency matrices to model spatial relationships between vehicles and utilizes Vision Transformer to extract global contextual features from multi-vehicle historical trajectories. An affine full attention mechanism is designed to fuse temporal multi-head attention and spatial graph attention networks, effectively capturing spatiotemporal interaction patterns and long-term temporal dependencies. A bidirectional long short-term memory network is adopted to construct an encoder-decoder architecture for end-to-end trajectory prediction. Experimental results on the NGSIM US-101 and I-80 datasets demonstrate that BiFSTNet achieves an average root mean square error of 1.77 m on the US-101 dataset and 2.06 m on the I-80 dataset, showing performance improvements over baseline methods across all prediction time horizons.