摘要
针对因为忽略车辆运动状态而导致的车辆预测轨迹不准确的问题,提出了一种基于运动状态的轨迹预测模型Movement-DenseTNT。首先,对交通参与者的轨迹信息和地图信息以图神经网络的方法进行编码;其次,使用LSTM提取车辆的运动状态信息;然后,将场景编码信息与在可行驶区域内采样得到的候选轨迹终点集合通过注意力机制的方式进行信息融合,从而得到每个候选轨迹终点的概率值;最后,通过筛选得到最终的轨迹终点并进行轨迹补全,以此得到准确的轨迹预测结果。该模型在两个基准数据集上与九个基线模型进行了比较,实验结果显示,Movement-DenseTNT模型在四个常用评估指标上优于基线模型,验证了加入车辆运动信息可以有效提升车辆轨迹预测的精度。
Aiming at the problem of inaccurate vehicle prediction trajectory caused by ignoring the vehicle motion state,this paper proposed a trajectory prediction model based on the motion state named Movement-DenseTNT.Firstly,the trajectory information and map information of traffic participants were encoded using the graph neural network method.Secondly,it used LSTM to extract the motion state information of the vehicle.Then,it fused the scene coding information with the set of candidate trajectory endpoints sampled in the drivable area through the attention mechanism,so as to obtain the probability value of each candidate trajectory endpoint.Finally,by filtering out the final trajectory endpoint and completing the trajectory,it obtained an accurate trajectory.The model was compared with nine baseline models on two benchmark datasets.The experimental results show that the Movement-DenseTNT model is superior to the baseline model in four commonly used metrics.It verifies that adding vehicle motion states can effectively improve the accuracy of vehicle trajectory prediction.
作者
顾一凡
莫磊
Gu Yifan;Mo Lei(School of Automation,Southeast University,Nanjing 210096,China)
出处
《计算机应用研究》
北大核心
2025年第4期1080-1084,共5页
Application Research of Computers
基金
江苏省自然科学基金面上项目(SBK20242028)
国家重点研发青年科学家资助项目(2022YFF0902800)。
关键词
轨迹预测
运动状态
注意力机制
图神经网络
trajectory prediction
motion state
attention mechanism
graph neural network(GNN)
作者简介
顾一凡(2000-),男,上海人,硕士研究生,CCF会员,主要研究方向为轨迹预测;通信作者:莫磊(1985-),男,广东中山人,副教授,博士,主要研究方向为嵌入式系统软硬件综合与智能计算(lmo@seu.edu.cn).
