Skip to Main content Skip to Navigation

Vision pour la robotique en environnement naturel

Abstract : In order to execute missions at a high level of abstraction, such as "go to - (landmark)", an autonomous mobile robot must be able to perceive its environment using its on board sensors (laser range finder, video camera...). Data issued form depth sensors offers a usefull description on ground geometry and object surfaces which is essential for robot navigation. Meanwhile, it does not permit to obtain a sufficient description of the environment : multiple objects may have the same shape, the ground may be unfit even if it is flat... ` This thesis deals with the development of the use of a video camera in order to add usefull information for both localization and navigation of a mobile robot roving in natural environments. The current needs for the realization of these two tasks offers two directions. The first concerns the three-dimensional information obtaining using a stereoscopic system ; the second is the addition of knowledge about environment by object nature identification. Stereovision is a well-known method to obtain depth information. We have implemented a stereocorrelation algorithm : a well-adapted technique for textured scenes. The basic algorithm was improved so as to take into account both the execution speed constraints and the correlation quality. When the robot executes a mission in an outdoor environment, the ground knowledge can be improved by adding information other than geometrical or topological, such as color or texture. We have proposed and implemented as a complementary representation, the nominally model of regions which indicates each region nature in an image. Firstly, a segmentation algorithm provides a synthetic description of the scene : the method combines both region growing method and clustering technique (based on general histogram shapes of each colorimetric component). Regions issued from the segmentation stage are then characterized and afterwards identified in order to obtain their nature (grass, rocks, ground...). Their characterization is obtained from their color and texture. The developed texture operators are based on contrast lines in the considered region (density, curve,...). Finally, a probabilistic method is used to determine the nature of the current elements in the environment.
Document type :
Complete list of metadata
Contributor : Emilie Marchand <>
Submitted on : Tuesday, April 3, 2007 - 4:20:03 PM
Last modification on : Friday, January 10, 2020 - 9:08:08 PM
Long-term archiving on: : Friday, May 13, 2011 - 9:27:53 PM


  • HAL Id : tel-00139846, version 1


Patricia Lasserre. Vision pour la robotique en environnement naturel. Automatique / Robotique. Université Paul Sabatier - Toulouse III, 1996. Français. ⟨tel-00139846⟩



Record views


Files downloads