GA-FPFH: A Global-Prior Augmented Fast Point Feature Histogram for Robust LiDAR SLAM Point Cloud Registration

Abstract

Backpack and handheld LiDAR simultaneous localization and mapping (SLAM) systems have become an important solution for large-scale 3D data acquisition. Since Global Navigation Satellite System (GNSS) positioning is not always available in many LiDAR SLAM systems, point clouds acquired from different surveying projects or devices are represented in independent local coordinate systems and require coarse registration to fuse all data into a unified coordinate system. Existing coarse registration approaches based on local feature descriptors often depend on locally estimated surface normals or reference directions, whose repeatability can be affected by measurement noise and non-uniform sampling. To address this issue, this paper proposes a Global-Prior Augmented Fast Point Feature Histogram (GA-FPFH) descriptor. The proposed method constructs a Z-axis-augmented local reference frame (Z-LRF) using the gravity-aligned vertical direction provided by the SLAM system. Three new geometric components are proposed based on the Z-LRF and combined with conventional FPFH features to form a six-component and 66-dimensional descriptor. Experiments on 12 real-world point-cloud pairs show that GA-FPFH increases the inlier ratio by 26.9–148.9% and, across five registration algorithms, reduces the rotation error, translation error, and RMSE by 31.5–87.9%, 29.6–97.8%, and 42.4–97.6%, respectively, while increasing the overall registration success rate from 71.7% to 88.3%. The significant error reductions are partly attributable to the higher registration success rate and fewer failure cases.