G-PROBE boosts LiDAR localization 18x under narrow FOV

Jinseop Lee’s G-PROBE addresses a critical gap in 3D point cloud localization: robustness to asymmetric or limited fields of view (FOV). Traditional methods assume dense, symmetric sensor coverage, but G-PROBE removes this constraint by introducing a virtual sensor decomposition that standardizes processing across narrow-FOV, panoramic, or multi-sensor setups. The front-end leverages cross-FOV branch ensembles to encode heading hypotheses, while a novel gamma-SGRT algorithm suppresses heading aliasing—becoming inert at full 360° coverage—without requiring tuning.

The back-end, CG-GICP, refines coarse GICP estimates by focusing only on high-certainty co-observed points, as identified by a bird’s-eye-view certainty map derived from front-end scoring. This certainty coupling directly links descriptor evaluation to 6-DoF pose estimation, eliminating the need for external verification modules. In benchmarks spanning five LiDAR datasets (including mechanical, solid-state, and FMCW sensors), G-PROBE achieved the highest multi-session F1 scores among learning-free methods and remained competitive in single-session panoramic settings. Notably, it maintained 54% Recall@1 under extreme FOV asymmetry (360°→60°), a 18x improvement over the strongest learning-free baseline.