基于动态图注意力的车辆轨迹预测研究

Research on Vehicle Trajectory Prediction Based on Dynamic Graph Attention

针对目前轨迹预测研究中交互建模方法使用的图注意力网络(GAT)为静态注意力,无法有效捕捉复杂道路场景中车辆间交互的问题,提出了一种基于编码器-解码器架构的动态图注意力网络(ED-DGAT)预测高速公路环境中运动车辆的未来轨迹。编码模块使用动态图注意力机制学习场景中车辆间的空间交互,采用状态简化动态图注意力网络建模解码阶段车辆运动的相互依赖,最后使用NGSIM数据集评估所提出的模型,并与长短时记忆(LSTM)、联合社交池化与长短时记忆(S-LSTM)、联合卷积社交池化与长短时记忆(CS-LSTM)算法模型进行对比分析,结果表明,预测轨迹的均方根误差(RMSE)降低了25%,且模型的推理速度为CS-LSTM模型的2.61倍。

In current research on vehicle trajectory prediction,the existing Graph Attention Network(GAT),which is based on a static attention mechanism,fails to effectively capture interactions between vehicles in complex road conditions.To address this issue,this paper proposed an Encoder-Decoder Dynamic Graph Attention Network(ED-DGAT)to predict future trajectories of highway vehicles.In this model,the encoding module incorporates a dynamic graph attention mechanism to learn spatial interactions among vehicles.Simultaneously,a simplified dynamic graph attention network is adopted to model the interdependencies of vehicle movements during the decoding phase.This paper evaluated the proposed algorithm using the NGSIM dataset and conducted comparative analysis with other models such as LSTM,SocialLSTM(S-LSTM),and CS-LSTM.The results show that the Root Mean Squared Error(RMSE)of predicted trajectory has been reduced by 25%,and the inference speed is 2.61 times of the CS-LSTM model.