This article presents a novel approach to real-time motion prediction in autonomous driving systems. The authors introduce the K-nearest neighbor attention with relative pose encoding (KNARPE), and the Heterogeneous Polyline Transformer with Relative pose encoding (HPTR), aiming to improve efficiency and scalability in dense traffic scenarios. The system shares contexts among agents and reuses unchanged contexts, significantly reducing computational overhead. The method performs on par with state-of-the-art agent-centric methods while maintaining efficiency. Tests on Waymo and Argoverse-2 datasets show superior performance among end-to-end methods without expensive post-processing or model ensembling.
Publication date: 19 Oct 2023
Project Page: https://github.com/zhejz/HPTR
Paper: https://arxiv.org/pdf/2310.12970
