Skip to Main content Skip to Navigation
Theses

Exploration visuelle d'environnement intérieur par détection et modélisation d'objets saillants

Abstract : A robot companion has to understand a domotic environment in order to execute requests like « Search a glass in the kitchen » with a high level of autonomy. So the robot must acquire several representations adapted to the tasks to be executed. This thesis proposes an on line learning method of an environment model expressed as a set of local structures described by appearance-based characteristics, and possibly named by a tutor. Such descriptions could be used in order to define a topological area (e.g. the kitchen) by a set of local structures or objects that could be found here (e.g. glasses, fridge, pans&). For the construction of such a representation, it is proposed a cognitive mehod, based on attentive and preattentive visual processes, acquiring images from a multifocal sensor. The preattentive process aims at detect interest regions, that could contain discriminant visual information ; based on the saliency concept proposed initially bu Itti and Koch, interest regions are extracted from a saliency map, built from images acquired by a short lens or panoramic camera (large view field). Such a region is then tracked on several successive images acquired while the robot is moving, so that the size and the 3D position of the corresponding local structure could be coarsely estimated. Then the attentive process exploits attentionam mechanisms in order to be focused successively on each interest region : it aims to characterize each local structure by an appearance-based model defined by an associative memort views-patches-aspects. Salient scaled patches or SIFT features are extracted from every image. After this exploration step is over, the robot tutor could annotated the model, segmenting local structures in objects, naming objects and grouping them in areas (kitchen&). Then, the robot exploits this environment model for the recognition and the coarse localization of objects and areas.
Document type :
Theses
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-00289380
Contributor : Arlette Evrard <>
Submitted on : Friday, June 20, 2008 - 2:33:23 PM
Last modification on : Monday, October 19, 2020 - 11:05:52 AM
Long-term archiving on: : Friday, September 28, 2012 - 4:20:25 PM

Identifiers

  • HAL Id : tel-00289380, version 1

Citation

Maxime Cottret. Exploration visuelle d'environnement intérieur par détection et modélisation d'objets saillants. Automatique / Robotique. Institut National Polytechnique de Toulouse - INPT, 2007. Français. ⟨tel-00289380⟩

Share

Metrics

Record views

300

Files downloads

983