Reproduce “Online Navigation Refinement: Achieving Lane-Level Guidance by Associating Standard-Definition and Online Perception Maps” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiaxu Wan, Xu Wang, Mengwei Xie, Xinyuan Chang, Xinran Liu, Zheng Pan, Mu Xu, Hong Zhang and 2 more
An approach for online navigation refinement combines global standard-definition maps with real-time online perception maps to deliver precise lane-level guidance. The Map Association Transformer (MAT) model, evaluated on a new dataset, achieves 78.2% NR-F1$_{50:95}$ and 34 ms latency, demonstrating robustness across diverse noisy perception inputs from various generators.
Key insights
- Effectively combining global, static topological information from SD maps with dynamic, real-time geometric details from OP maps enables precise and up-to-date lane-level navigation without relying on costly HD maps.
- A transformer architecture equipped with both local path-aware and global spatial attention mechanisms can robustly handle the significant spatial and semantic heterogeneity between road-level and lane-level map data.
- Leveraging SD map priors provides resilience against various real-world challenges, including noise, spatial misalignments, missing data in SD maps, and adverse environmental conditions affecting OP map quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.