Fast Point Feature Histograms

A hand-crafted feature that describes each point’s local shape using a histogram of normal-vector geometric relationships in its neighborhood.

FPFH was proposed by Radu Rusu, Nico Blodow, and Michael Beetz at ICRA 2009, as a faster version of PFH (Point Feature Histograms). The process estimates a normal vector for every point, then computes several angular relationships between it and its neighboring points, summarized into a histogram as the descriptor. PFH computes relationships between every pair of points in a neighborhood, with complexity O(nk²); FPFH only computes relationships between the center point and each neighbor, then combines neighbors’ results weighted by distance, bringing complexity down to O(nk) with little loss of discriminative power. The default implementations in both PCL and Open3D are 33-dimensional. It requires no training, and its most typical use is global point cloud registration: FPFH first finds correspondences between two point clouds, RANSAC solves for a rough pose, and ICP then refines it.

ExampleOpen3D’s global registration tutorial first voxel-downsamples two point clouds and estimates normals, computes 33-dimensional FPFH features, uses RANSAC to solve for an initial pose, and then refines it with ICP.

Also called

FPFH, FPFH Descriptor

Related

Point Cloud Registration · Surface Normal Estimation · Iterative Closest Point · Random Sample Consensus · Point Cloud Library (PCL) · Open3D

Sources

PCL Tutorials: Fast Point Feature Histograms (FPFH) descriptors
Open3D: Global registration