A comparison study of algorithms for surface normal determination based on point cloud data

Abstract

Robot applications in manufacturing of aircraft sheet metal parts require real-time methods for determining the surface normal using a digitized point data set measured by a 3D laser scanner. For this reason, six archived surface normal algorithms are compared. Though using different weights, these methods are all set to determine the surface normal at a given point by averaging the surface normals of the adjacent facets. In this paper, a comparison study is designed with a nearest neighboring method searching for adjacent facets, along with the introduction of a dynamic sampling method to investigate the effect of the resolution of a data set on the accuracy of surface normal determination. Three performance indices are proposed including the total number of final data points, the number of times of up-sampling and the total computing time. Three geometric models are considered including a sphere representing an aircraft cockpit, a cylinder representing a fuselage, and an ellipsoid representing a wing. The laser scanner error is modeled by a log-normal distribution. While all the six methods can generate satisfactory results in error-free case, the simulation results indicate that in error case MWE (mean weighted equally) and MWAAT (mean weighted by areas of adjacent triangles) are not favorable while the other four methods exhibit no obvious difference.