Thèse soutenue

Auto-qualification de données géographiques 3D par appariement multi-image et classification supervisée : application au bâti en milieu urbain dense

FR  |  
EN
Auteur / Autrice : Laurence Boudet
Direction : Nicolas PaparoditisMarc Pierrot-Deseilligny
Type : Thèse de doctorat
Discipline(s) : Informatique
Date : Soutenance en 2007
Etablissement(s) : Université de Marne-la-Vallée (1991-2019)

Mots clés

FR

Mots clés contrôlés

Résumé

FR  |  
EN

Despite advances in automatic scene reconstruction field and the current technological transfer from research to production, modeling errors remain due to building shape complexity and diversity in urban areas. Produced data have to be checked, but only human verification of individualized data handles this task. This significantly reduces the interest of automation in scene reconstruction. In this work, we propose a new option that aims at automating 3D data quality diagnosis with the only available external data : the aerial images. This option takes place in a semi-automatic framework, where a human operator will be able to verify subsequently all the data that have not been validated by the process. The process is composed of three steps : observation extraction from the images, their comparison to the 3D data and a decision step. In this work, we have chosen the multi-image option that leads to data evaluation in the 3D space. First, we have been interested in matching several images from textures, from radiometries as well as from structures. 3D observations of diverse kinds have been extracted in order to verify geometrical and structural properties of 3D data. Furthermore, method genericity is ensured by the number of observations. Second, we have proposed to use a supervised classification method to cope with the decision process. This method enables to gather knowlegde on data quality as well as to solve the problem, that is to say to predict the quality of a new 3D data. Considering the performance constraints linked to an automatic validation system, we have then proposed to apply robust decision rules. These rules are in keeping with the traffic light paradigm and are very selective about the validation. Finally, we have applied the method to real buiding data in dense urban areas and performed many evaluations