基于潜在特征的时空图卷积网络轨迹预测方法

Trajectory Prediction Method with Spatial-Temporal Graph Convolutional Network Based on Latent Features

为提高车辆轨迹预测精度,提出一种基于潜在特征的时空图卷积网络轨迹预测方法CRSTGCN。首先,该方法特别添加了一个时间上更早、更长的历史轨迹作为输入,并基于该输入建立了潜在特征编码层。然后,CR-STGCN将该潜在特征编码层编码的潜在特征与时空图卷积编码的机动性与动力性特征拼接融合,并采用两层门控循环单元(Gate Recurrent Unit,GRU)解码出预测轨迹。最后,将采用时空图卷积编码和两层GRU解码的预测轨迹模型STGCN与CR-STGCN在NGSIM数据集上进行对比。结果表明,CR-STGCN在不同机动类型、交通密度场景下的预测精度均优于STGCN,证明了这一方法应用于车辆轨迹预测的有效性,为轨迹预测特征选取提供了新思路。

In order to increase the accuracy of vehicle trajectory prediction,this paper proposed a trajectory prediction method with Spatial-Temporal Graph Convolutional Network:CR-STGCN,which was based on latent features.Firstly,an earlier and longer historical trajectory was added specially as input,and a latent feature encoding layer was established based on the input.Next,CR-STGCN concatenated and fused the latent features encoded by the encoding layer of the latent features with the mobility and dynamic features encoded by the spatial-temporal graph convolutional network,and then the predicted trajectory was decoded using a two-layer gate loop unit GRU.Finally,the predicted trajectory model STGCN using spatial-temporal convolutional network encoding and two-layer GRU decoding was compared with CRSTGCN on the NGSIM dataset.The results show that the prediction accuracy of CR-STGCN is higher than STGCN in different types of maneuvers and traffic density scenarios,demonstrating the effectiveness of this method in vehicle trajectory prediction and providing a new approach for feature selection in trajectory prediction.