A Lightweight Multi-Scale Hierarchical Network for Multi-Modal Trajectory Prediction in Autonomous Driving

Accurately predicting the trajectories of surrounding vehicles is critical for safe and efficient autonomous driving. Although map-based models have demonstrated strong performance, their reliance on high-definition maps limits their applicability to scenarios where such maps are unavailable or incomplete. To tackle this issue, a lightweight map-free trajectory prediction model is proposed. This model is composed of two main components, a spatial-temporal trajectory encoder module (STEM) and a spatial-temporal multi-agent interaction module (SMIM). The STEM component is used to extract spatial and temporal features by combining attention mechanism. The SMIM component uses Subequivariant Graph Neural Networks (SGNN) for spatial interactions and a Mamba-based state space module for temporal interactions. Moreover, the model generates multimodal trajectories through a decoder inspired by Bernstein basis polynomials. Extensive experiments conducted on the Argoverse1 dataset demonstrate that the proposed model can achieve competitive performance compared with map-based methods. Notably, it requires fewer parameters and eliminates the reliance on map data. This characteristic makes our approach highly suitable for autonomous driving scenarios with unavailable or incomplete map data.