Shape Decomposition Using Structured Light Vision

Abstract

Vision systems are becoming an integral part of many automated manufacturing processes involving localisation and identification of 3D shapes. Active vision offers a practical solution to such problems. In this paper, we present a new recursive method to extract quadratic surfaces when scanning a scene with a plane of laser light.

At each scanning step, a statistical analysis of the intensity arrangement in the image is used for extracting the stripe skeleton and for locating its breaking points which correspond to occluding boundaries between faces. Then, a fuzzy analysis of the stripe fragments allows us to locate curvature discontinuities. So, each stripe is segmented into several parts which can be approximated by second order curves. These parts are tracked in consecutive images and matched in order to create regions. The decision process is controlled by using fuzzy values associated to likeness criteria between the matched stripes. The segmentation problem being solved, the 3D shape recovery can be performed via triangulation for computing 3D points and least square minimization for identifying quadratic models of the faces.

We present the segmentation algorithms and some results obtained with real scenes consisting of multiple objects of arbitrary shapes. They show that an efficient shape decomposition may be obtained with few-constrained environments including planar or curved shapes.