Modeling spatio-temporal interactions for vehicle trajectory prediction based on graph representation learning

Accurately predicting vehicle trajectories is essential for safe and efficient operations of autonomous driving cars. In this paper, we propose a long short-term memory (LSTM) encoder-decoder model along with a graph representation learning module, where LSTM is adopted to deal with temporal sequence features and a graph representation learning module is used for precisely learning the spatial interactions between vehicles. By leveraging this representation learning module based on an attention aggregator, the developed model can learn the attention paid by drivers to the vehicles in their local neighborhood and aggregate forward traffic flow information, thereby significantly improving the long-term prediction accuracy and the interpretability of the model. In the conducted experiments, the publicly available Next Generation Simulation (NGSIM) US-101 and I–80 datasets are used to train and evaluate our model. The results indicate that our spatial-temporal framework achieves noticeably better performance than the state-of-the-art approaches. Specifically, we reduce the root mean square error (RMSE) by 17% according to quantitative evaluation.