Reconnaissance d'Objets Polyédriques à partir d'une image vidéo pour la téléopération

Abstract

Our laboratory is developing an interface system for Telerobotics MCIT. The aim is to provide visual aid and control of the site for the operator. The visual aid consists in updating and superimposing a 3DDB to a video image. In this thesis, a polyhedra recognition system from a single 2D image has been developed into MCIT to solve the last problem. Image processing software based on object oriented library has been developed, within which an improvement of the Hough transform was achieved to better extract image line segments. Perceptual organization is also used to provide a 2D model of the image. Two matching methods have been applied: the graph method which gives a minimal number of hypotheses using projective invariants but, it fails when the image processing is of poor quality. In this case, we apply the geometric hashing method, which always provides a solution. We have developed two aspect graph extraction methods applicable on polyhedra. Each one is used by one of the above-mentioned matching methods. For object localization, we apply a hybrid technique that makes use of three well-known reconstruction methods, which gives a better precision. The automatic camera calibration using a 4dof robot has been developed to reduce system errors.