Share Email Print
cover

Proceedings Paper

Neural-network calibration of a multiple-line laser-camera range sensor for 3D surface-geometry measurement
Author(s): Chris Yu-Liang Liu; Jonathan Kofman
Format Member Price Non-Member Price
PDF $14.40 $18.00

Paper Abstract

Single-line laser-camera range sensors require scanning over the object surface to measure three-dimensional (3D) surface geometry. Full-field 3D surface measurement techniques typically require more than one pattern to be projected and captured by camera. This paper presents a method to calibrate a multiple-line laser-camera range sensor using an artificial neural network (NN) to enable capture of full-field 3D surface geometry using a single projected pattern. The range sensor projects nineteen laser lines onto a surface. During calibration, points in 2D images are extracted from the intersections of nineteen laser profiles and horizontal lines marked on a calibration plate, for several calibration plate positions. A mapping of 2D image coordinates to 3D object coordinates is performed separately for each laser-line projection using a multi-layer perceptron (MLP) neural network. Experiments using different NN configurations found a network with two hidden layers of 43 nodes per layer using the sigmoidal activation function to generate the lowest 3D reconstruction errors. Errors were consistent errors over all calibration positions. Calibration with an acceptable error for many applications can be achieved without knowledge of the camera pose. The fast 3D reconstruction by the trained system may permit low resolution full-field 3D surface-geometry measurement in real-time.

Paper Details

Date Published: 17 November 2008
PDF: 6 pages
Proc. SPIE 7266, Optomechatronic Technologies 2008, 72661J (17 November 2008); doi: 10.1117/12.817411
Show Author Affiliations
Chris Yu-Liang Liu, Univ. of Waterloo (Canada)
Jonathan Kofman, Univ. of Waterloo (Canada)


Published in SPIE Proceedings Vol. 7266:
Optomechatronic Technologies 2008
John T. Wen; Dalibor Hodko; Yukitoshi Otani; Jonathan Kofman; Okyay Kaynak, Editor(s)

© SPIE. Terms of Use
Back to Top