Skip to Main content Skip to Navigation
Theses

Quelques contributions des invariants projectifs à la vision par ordinateur

Abstract : One of the main goals in computer vision is to infer three-dimensional characteristics of some observed objets by analyzing one or several images of them. This allows either to determine the shape and position of these objects or to recognize them. Classical positioning methods depend on a previous calibration of the cameras. The calibration process is tedious and unstable, and sometimes even impossible, as it is the case with a moving camera. We show that using the properties of projective geometry, one can avoid explicit calibration and obtain a relative positioning of the observed objects. We validate our method by doing some experiments on real images of simple polyhedric scenes, evaluating also its precision. The development of techniques that use projective invariants for positioning or recognition leads us to study the stability of projective invariants when computed from noisy images. A theoretical study allows us to propose a similarity measure for projective invariants, and some methods to identify and eliminate unstable configurations. We then consider different characterizations of five coplanar points with projective invariants and we compare their performance in a recognition process using simulated noisy data.
Document type :
Theses
Complete list of metadatas

Cited literature [33 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00005137
Contributor : Thèses Imag <>
Submitted on : Thursday, February 26, 2004 - 5:01:28 PM
Last modification on : Friday, November 6, 2020 - 4:39:30 AM
Long-term archiving on: : Thursday, September 13, 2012 - 12:00:38 PM

Identifiers

  • HAL Id : tel-00005137, version 1

Collections

UGA

Citation

Luce Morin. Quelques contributions des invariants projectifs à la vision par ordinateur. Interface homme-machine [cs.HC]. Institut National Polytechnique de Grenoble - INPG, 1993. Français. ⟨tel-00005137⟩

Share

Metrics

Record views

318

Files downloads

469