BACK

Modélisation et estimation du mouvement et de la structure dans une séquence d'images à partir d'indices géométriques épars, avec application à la post-production audio-visuelle

P. L. Bazin

PhD dissertation
(note: entirely written in French)

Abstract

Shape and motion recovery is a central problem of computer vision, that can be formulated in many different ways. For post-production applications, the main difficulty is to recover precisely the 3D structure and the path of the camera in long image sequences. Our approach states the problem as a parametric estimation problem, that allow to handle heterogeneous informations on the system and uncertainties in a Bayesian context. Using non-linear models designed by geometric reduction, the scene structure can integrate various geometric shapes like line segments, rectangles or 3D corners, along with geometric relationships. Similarly, the camera motion is modeled on the entire sequence using polynomial curves, which degree is set by model selection. The geometric elements of the scene are tracked along the sequence with a coupled correlation/matching technique. The parameters of the scene and camera models are then estimated from the data extracted during the tracking in a sequential fashion. The performances of the algorithm are characterised with reconstruction results on various simulated and real sequences, and the method is succesfully applied for augmented reality effects.


Document intégral (postscript)

Table des Matières

  1. Le problème de Structure et Mouvement (Chapter (postscript))

  2. Modèles de scènes géométriques (Chapter (postscript))

  3. Modélisation de la caméra et de son mouvement (Chapter (postscript))

  4. Poursuite d'indices épars dans les séquences d'images (Chapter (postscript))

  5. Estimation statistique du mouvement et de la structure (Chapter (postscript))

  6. Application logicielle et résultats (Chapter (postscript))

  7. Extensions (Chapter (postscript))

  8. Annexes (Chapter (postscript))

BACK