HSTR: Hierarchical Scene Transformer for Multi-agent Trajectory Prediction

Trajectory prediction plays a pivotal role in the autonomous driving systems. Existing methods generally employ agent-centric or scene-centric approaches to represent driving scenarios. However, these methods introduce significant redundant computations or pose losses, resulting in suboptimal prediction efficiency and accuracy. To tackle these problems, a novel multi-target trajectory prediction model, named Hierarchical Scene TRansformer (HSTR), is introduced. The driving scene is decomposed into two independent components by HSTR: global and local. In the global part, global interaction information is established and shared among all predicted agents, thereby reducing redundant computations. In the local part, an individual reference frame is established for each vehicle to eliminate the impact of pose variations and extract temporal features. Moreover, an adaptive anchor point generation method is proposed to address the challenge of capturing future modalities for vehicles. This method dynamically generates corresponding anchor points based on different driving scenarios to guide the prediction of trajectories across various modalities. The model performance is verified on the argoverse1 and argoverse2 datasets, and the experimental results demonstrate that competitive performance is achieved by HSTR in terms of efficiency and precision compared to the state-of-the-art methods.