Scan corner of context: a novel corner feature rotation-invariant descriptor for loop closure detection from 3D LiDAR data

Loop closure detection is essential in SLAM for identifying revisited locations and establishing inter-frame pose constraints, which minimizes error accumulation and ensures accurate localization and mapping. Current LiDAR-based loop closure detection (LCD) methods are effective in outdoor environments but often perform poorly indoors. We propose a new rotation-invariant descriptor for corner point features in loop closure detection, called the context scan corner algorithm. First, we optimize the original scan context descriptor through coarse matching of corner points and introduce a corner-based descriptor named “scan corner of context” (SCC) for matching. Second, we achieve rotation invariance by using principal component analysis (PCA) to estimate the angle of change between two frames of point clouds. Finally, we extensively evaluate our method with both indoor and outdoor datasets. Experimental results indicate that the scan corner of context provides high accuracy and strong recognition capabilities in edge-rich scenarios. The SCC achieves higher F1-scores than the next best algorithm by 0.1, 0.1, 0.02, and 0.07 on the KITTI_08, Indoor, Hall_05, and NCLT datasets, respectively.