SEAM is a LiDAR lifelong mapping framework that anchors its evidence to submaps instead of one anchor for the whole session. Because evidence is reprojected per submap, it stays valid when a later session corrects the trajectory, so dynamic-object removal and change detection need not be recomputed from scratch. It suppresses geometrically weak inter-session loop edges with a DOP-based confidence and uses a direction-aware voxel evidence model to separate moving objects from real change, tested on a construction site and a long-term multi-session dataset.
Relevance
Lifelong maps break when the past trajectory is corrected, and this is a direct fix. Read for long-term autonomy and dynamic scenes; rejecting weak inter-session loops is also a robust estimation question.
Links
- arXiv
- Subfield: Semantic mapping, Mapping and localisation