Skip to Main content Skip to Navigation
Theses

Estimation de mouvement et segmentation
Partie I : Estimation de mouvement par ondelettes spatio-temporelles adaptées au mouvement.
Partie II : Segmentation et estimation de mouvement par modèles de Markov cachés et approche bayésienne dans les domaines direct et ondelette.

Abstract : The first part of this thesis presents a new vision of the motion estimation problem, and hence of the compression of video sequences. On one
hand, we have chosen to investigate motion estimation from redundant wavelet families tuned to different kind of transformations and, in particular,
to speed. These families, not well known, have already been studied in the framework of target tracking. On the other hand, today video standards,
like MPEG4, are supposed to realize the compression in an object-based approach, but still compute raw motion vectors on ``blocks''. We thus
implemented these wavelet families because 1) they are built to perform motion parameter quantization on several kinds of motions
(rotation, speed, acceleration) and 2) we think that an approach of motion estimation based on the trajectory identification of objects
motions in a scene is an interesting solution for future compression methods. We are convinced that motion analysis, and
understanding, is a way of reaching powerful ``contextual'' compression methods.


The second part introduces two new methods and algorithms of unsupervised classification and segmentation in a Bayesian approach. The first one,
dedicated to the reduction of computation times in the segmentation of video sequences, is based on an iterative, simple, implementation of the
segmentation. It also enabled us to set up a motion estimation based on objects segmentation. The second is aimed at reducing the segmentation times,
for images, by performing the segmentation in the wavelet domain. Both algorithms are based on a Bayesian estimation approach with a Potts-Markov
random field (PMRF) model for the labels of the pixels, in the direct domain, and for the wavelet coefficients. It also uses an iterative MCMC
(Markov Chain Monte-Carlo) algorithm based on a Gibbs sampler. The initial PMRF model, in the direct domain, works with a first order neighboring. We
have developed the PMRF model to tune it to the priviledged orientations of the wavelet subbands. These realizations provide, to our knowledge, new
approaches to unsupervised segmentation methods.
Complete list of metadatas

Cited literature [178 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00011310
Contributor : Patrice Brault <>
Submitted on : Friday, January 6, 2006 - 12:34:45 PM
Last modification on : Wednesday, September 16, 2020 - 4:55:48 PM
Long-term archiving on: : Friday, September 14, 2012 - 4:55:19 PM

Identifiers

  • HAL Id : tel-00011310, version 1

Collections

Citation

Patrice Brault. Estimation de mouvement et segmentation
Partie I : Estimation de mouvement par ondelettes spatio-temporelles adaptées au mouvement.
Partie II : Segmentation et estimation de mouvement par modèles de Markov cachés et approche bayésienne dans les domaines direct et ondelette.. Traitement du signal et de l'image [eess.SP]. Université Paris Sud - Paris XI, 2005. Français. ⟨tel-00011310⟩

Share

Metrics

Record views

622

Files downloads

2575