Dynamic-learning spatial-temporal Transformer network for vehicular trajectory prediction at urban intersections

Abstract

Forecasting vehicles' future motion is crucial for real-world applications such as the navigation of autonomous vehicles and feasibility of safety systems based on the Internet of Vehicles (IoV). Vehicular trajectory prediction at urban intersections remains challenging due to the difficulty in modeling temporal dependencies and spatial interactions among traffic agents. This paper proposes a dynamic-learning spatial-temporal Transformer network (DSTTN) with domain adaptation training methods based on two modules for two-dimensional vehicular trajectory prediction at urban intersections. The first module is trajectory maneuver characterization (TMC) which captures latent driving maneuver features and divides trajectory data into different categories with different distributions, which can be regarded as different domains in transfer learning (TL). The second module is trajectory distribution matching (TDM) which adopts a novel spatial-temporal Transformer network with distribution matching loss to dynamically learn domain-invariant maneuvers and achieve accurate trajectory prediction. Experiments and ablation studies collected at two unsignalized urban intersections of the inD dataset first validate the interpretability, transferability, and high prediction accuracy of the proposed DSTTN. The results show that the TMC module has maneuver-representational ability. The proposed STTN achieves good and stable prediction accuracy compared with other baselines, including the state-of-the-art deep learning model, while DSTTN with the TMC and TDM modules further improves accuracy. Additional experiments have been conducted under 70 randomly-selected urban intersection scenarios of the Waymo dataset to validate the good prediction accuracy of the proposed method. The formulation of DSTTN offers new ideas for dividing trajectory data into different domains and general domain adaptation to vehicle trajectory prediction under urban road environments.