Vision 3D non calibrée : contributions à la reconstruction projective et étude des mouvements critiques pour l'auto-calibrage

摘要

This thesis is about threedimensional reconstruction of objects from images taken by cameras. The classical scheme is based on an off-line calibration of the cameras. Very often, this is not available and, in general, dynamic applications need a frequent update of the calibration. Thus it is clearly desirable to get rid off the requirement of performing an off-line calibration. In this thesis, we concentrate on the one hand on methods for uncalibrated reconstruction and on the other hand on the concept of on-line or self-calibration. We try to get threedimensional measurements even from images taken with uncalibrated cameras. The obtained information is not of a metric nature, but still useful for tasks such as object recognition or the visually guided control of robot motions. We have developed several methods for uncalibrated threedimensional reconstruction. Self-calibration consists of calibrating a camera only from images of unknown objects. Images taken in course of an application can thus be used simultaneously for the application itself and for camera calibration. Our major contribution in this domain is a study of degeneracy conditions that turn out to be a very common problem in practice. In concrete terms, we identify all types of camera motion that prevent self-calibration.