Segmentation of Noisy Laser-Scanner Generated Meshes With Piecewise Polynomial Approximations

Abstract

Laser scanners offer a fast and simple way of collecting large amounts of geometric data from real-world objects. Although this aspect makes them attractive for design and reverse engineering, the laser-scanner data is often noisy and not partitioned into meaningful surfaces. A good partitioning, or segmentation, of the scanner data has uses including feature detection, surface boundary generation, surface fitting, and surface reconstruction. This paper presents a method for segmenting noisy three-dimensional surface meshes created from laser-scanned data into distinct regions closely approximated by explicit surfaces. The algorithm first estimates mesh curvatures and noise levels and then uses the curvature data to construct seed regions around each vertex. If a seed region meets certain criteria, it is assigned a region number and is grown into a set of connected vertices approximated by a bicubic polynomial surface. All the vertices in a region are within known distance and surface normal tolerances from their underlying surface approximations. The algorithm works on noisy or smooth data and requires little or no user interaction. We demonstrate the effectiveness of the segmentation on real-world examples.