Skip to Main content Skip to Navigation
Theses

Approche parallèle pour l'estimation du flot optique par méthode variationnelle

Abstract : This thesis deals with optical flow estimation. It was made in a multidisciplinary context. In fact, this work is the resolution of a fundamental problem in the image processing area (optical flow) by using scientific computing methods (parallel algorithms) to solve an experimental application from fluids mechanics area (Partical Image Velocimetry). The optical flow is defined as the vector field discribing the apprent motion between two successive images. It is an important source of information, so that, it is used in different kind of applications. We propose here to apply some image processing tools to particle image velocimetry (PIV). After a brief evaluation of the local and global regularization of vector field used for optical flow estimation, we consider the CLG method of Bruhn et al. [2005b] which proposes to combine these two kinds of regularizations. We first study the implementation of the warping in the large displacement contest and we justify it in our modified CLG (MCLG) approach. We describe then two solvers. The first uses variational approaches to derive a bloc Picard-SOR solver only employing local linearisations. We also described a color labelling scheme allowing the parallelisation of this solver. The second approach rely on a variable splitting technique which introduces an auxiliary variable in the considered model and to split the problem into two sub-problems easier to solve. This method only involve independent operations, its parallelisation is then simple. A full evaluation of the developped algorithms is presented. We give particular attention to their parallel properties on different architectures (multicore CPU and GPU). We then study their precision using some classical image proccessing datas and also PIV datas.
Complete list of metadatas

Cited literature [85 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00713970
Contributor : Riadh Fezzani <>
Submitted on : Tuesday, July 3, 2012 - 10:48:46 AM
Last modification on : Monday, December 14, 2020 - 9:55:40 AM
Long-term archiving on: : Thursday, October 4, 2012 - 2:42:30 AM

Identifiers

  • HAL Id : tel-00713970, version 1

Citation

Riadh Fezzani. Approche parallèle pour l'estimation du flot optique par méthode variationnelle. Traitement des images [eess.IV]. Université Pierre et Marie Curie - Paris VI, 2011. Français. ⟨tel-00713970⟩

Share

Metrics

Record views

693

Files downloads

3782