Ph.D. Thesis (in French)

Abstract: This thesis presents several registration methods for augmented reality applications. We first present two applications which use 2D composition techniques (based on homography estimation or non-rigid image transformation). We then propose a 3D visual tracking approach that makes use of a point-based model of the object. The idea is to explain the motion of feature points matched from frame to frame in terms of the variation of the 3D pose parameters. For that we compute 2D motion fields induced by elementary 3D motions. We adapt the splatting technique developed by the computer graphics community to render the model and approximate the motion fields anywhere on the surface. To avoid the drift of the estimation process, we use a set of keyframes which poses are determined offline. Our implementation also exploits the graphic processing units (GPU) for its most computationally intensive parts.

The Jury was composed of: Pascal Guitton (INRIA Futurs, Univ. Bordeaux I, France), Jean-Marc Lavest (Univ. Clermont I, France), Vincent Lepetit (CVLab, EPFL), Ariel Choukroun (from the french startup FittingBox) and my advisors Alain Ayache, Jean-Pierre Jessel, Vincent Charvillat and Géraldine Morin.

Dissertation | Defense slides (PDF only) | Defense slides with embedded videos (154 Mo)

See also here for other communications in French.