Vehicle Trajectory Prediction based on LSTM and Graph Convolutional Networks

Abstract

In order to improve the prediction accuracy of the future trajectory of a vehicle under the influence of surrounding vehicles and the surrounding road environment, we propose a vehicle trajectory prediction model based on LSTM and graph convolutional network. Firstly, the historical vehicle trajectory features of the target vehicle and a certain range around it are extracted using a modified LSTM encoder. Additionally, we model inter-vehicle interactions through a self-attentive mechanism. Secondly, we construct a high-definition map by vectorizing the lane map and introduce the inflated convolution operator to aggregate topological relationships between lane nodes. We fuse the vehicle interaction information and lane node information using residual blocks that incorporate the graph attention mechanism, Generating predictions with Multi-modal predictor heads. Finally, We conducted experiments using the publicly available dataset Argoverse, and the results show that our predictive model outperforms other models in terms of comprehensive performance.