Improved Point Pair Feature based Cloud Registration on Visibility and Downsampling

Abstract

Point clouds has been increasingly used in computer vision tasks like 3D reconstructions and robotic perceptions, and point cloud registration plays a key role in those scenarios. PPF (Point Pair Feature) based voting matching scheme is a widely-used method for point clouds registration. Based on actual experiences of using PPF we proposed modifications from 2 aspects. The first is based on point-pair visibility during offline dictionary generation and we applied 2 steps of spherical dictionary generation and merging instead of brute-force traversal. The second came from high-frequency information loss (richness of surface normal distribution decreasing) during point cloud downsampling, and we proposed a curvature-aware voxel downsampling method instead of uniform downsampling. We demonstrated the effectiveness of the proposals above via several experiments.