A Road-Aware Neural Network for Multi-step Vehicle Trajectory Prediction

Multi-step vehicle trajectory prediction has been of great significance for location-based services, e.g., actionable advertising. Prior works focused on adopting pattern-matching techniques or HMM-based models, where the ability of accurate prediction is limited since patterns and features are mostly extracted from historical trajectories. However, these methods may become weak to multi-step trajectory prediction when new patterns appear or the previous trajectory is incomplete. In this paper, we propose a neural network model combining road-aware features to solve multi-step vehicle trajectory prediction task. We introduce a novel way of extracting road-aware features for vehicle trajectory, which consist of intra-road feature and inter-road feature extracted from road networks. The utilization of road-aware features helps to draw the latent patterns more accurately and enhances the prediction performances. Then we leverage LSTM units to build temporal dependencies on previous trajectory path and generate future trajectory. We conducted extensive experiments on two real-world datasets and demonstrated that our model achieved higher prediction accuracy compared with competitive trajectory prediction methods.