HGeo-TopoMap uses hierarchical geometric priors to boost topological mapping

Autonomous driving perception systems rely on topological maps—detailed road representations showing centerlines, traffic signs, and their connectivity—for path planning. A major hurdle is detecting centerline instances because real-world roads lack explicit centerline markings. To address this, Siyu Li and colleagues introduce HGeo-TopoMap, which incorporates hierarchical geometric priors from two sources: an explicit prior map (e.g., from inverse perspective mapping) and implicit spatial relations among road elements.

The method comprises two key modules. First, a geometric adaptive learning module discretely encodes semantic and spatial features from the road structure map, then applies a novel prior-mask attention mechanism to focus only on informative regions, ignoring noise. Second, a geometric consistency learning module enforces spatial alignment between centerline instances that share identical geometric orientations, using a geometry-aware decoder. Evaluated on the OpenLane-V2 dataset across centerline, lane segment, and robustness benchmarks, HGeo-TopoMap consistently outperforms existing baselines—especially in challenging conditions like occlusions or poor lighting. The authors will release source code and model weights publicly, enabling further research and real-world deployment.