Calibration, rectification et stereoscopie (French and English text)

This thesis concentrates on three aspects of computer vision, all related to 3D reconstruction from multiple images of a scene: camera calibration, rectification of stereoscopic images, and stereoscopic matching.; Camera calibration is a crucial problem in computer vision. It consists in computing the relative displacement between two cameras, as well as internal camera distortions, by using only images taken by these cameras. This thesis will present a new calibration method that does not depend on the availability of established corresponding points.; Once the calibration is obtained, it becomes possible to establish a full correspondence between two images using stereo analysis. The fact that most of these methods establish correspondence between horizontal epipolar lines implies that a prior rectification step is needed when the camera displacement is not horizontal.; We present here a new rectification method, known as cylindrical, that can handle arbitrary camera displacements while preserving constant the size of the resulting rectified images.; The third part of this thesis proposes a new stereoscopic algorithm for establishing correspondence. It handles directly two or more arbitrary views simultaneously and does not require prior rectification of the images. Based on the computation of maximum flow in a graph, this method is the first to solve efficiently and optimally the correspondence problem with a smoothing constraint without relying on the epipolar constraint. Also described is the extension of this method to Markov random fields, thus broadening its application to other fields than stereoscopy.