Quelques outils de géométrie différentielle pour la construction automatique de modèles CAO à partir d'images télémétriques

Abstract : In the industry, there is a need for CAD models of the as-built large-scale architectures. At the present, these models may be obtained with photogrammetry, a slow and costy technology based on taking photographs of the structures from multiple view. A recent technology known as laser range sensing is able to give directly dense images of 3D points scanned on the surfaces of objects. A dedicated software is then capable to build a CAD model fitted on the measure points. The main difficulty in building a CAD model from range images lays in the ability to segment the data sets into subsets corresponding each one to a unique geometric primitive, such as cylinder, torus, sphere, cone or plane. This problem being particularly difficult to overcome, the existing software require the help of the operator to interactively do the segmentation on the screen. The goal of this thesis was to explore the automation possibilities of this process. In a first step the study focused on pipes, which represent the main part of the scenes observed. These elements may be modeled with cylinders, torii and cones. The proposed approach consists in segmenting the pipes using the local centers of curvature of the surfaces. These centers of curvature draw lines in the 3D space, easy to segment afterwards and from which it is possible to go back to the original images. To compute the centers of curvature, it has been necessary to perform a theoretical study of the algorithm giving the principal curvatures on discrete surfaces, study which lead to an improvement of the algorithm from what can be found in the litterature, and to the definition of a noise optimality criterion. The algorithms have been tested on several industrial images. The study went up to the automatic CAD reconstruction of a small pipe part, validating the approach. The original goal of the thesis, automatic CAD segmentation, has therefore been reached. However, many research topics are still to be studied and new ones came out, about improving the algorithms or studying planes in the images for example.
Complete list of metadatas

Cited literature [17 references]  Display  Hide  Download

https://pastel.archives-ouvertes.fr/pastel-00727509
Contributor : François Goulette <>
Submitted on : Monday, September 3, 2012 - 5:27:05 PM
Last modification on : Monday, November 12, 2018 - 10:54:19 AM
Long-term archiving on: Tuesday, December 4, 2012 - 3:42:18 AM

Identifiers

  • HAL Id : pastel-00727509, version 1

Citation

François Goulette. Quelques outils de géométrie différentielle pour la construction automatique de modèles CAO à partir d'images télémétriques. Vision par ordinateur et reconnaissance de formes [cs.CV]. École Nationale Supérieure des Mines de Paris, 1997. Français. ⟨NNT : 1997ENMP0735⟩. ⟨pastel-00727509⟩

Share

Metrics

Record views

619

Files downloads

1917