Skip to Main content Skip to Navigation
Theses

Approche monodimensionnelle de la mise en correspondance stéréoscopique par corrélation - Application à la détection d'obstacles routiers

Abstract : In stereovision, a dense disparity map can be computed thanks to local methods that match 2D image neighbourhoods. In the context of 3D metrology, obtaining a dense disparity map is not mandatory, but high accuracy is required. Moreover, because of important local variations of disparity values due to high perspective conditions, we show that 2D methods do not reach a sucient level of accuracy. In this work, we propose a new matching method based on the analysis of several characteristics extracted from a set of similarity curves computed using 1D neighbourhoods. We compute condence values directly linked to accuracy and density properties of the disparity map. This condence values are used to tune a trade-o between density and accuracy of the disparity maps. The generic proposed method is evaluated by comparing its results with those computed with six similar local methods. In this way, we show its eciency, particularly with scenes having a strong perspective. Finally, we validate our 1D method in an obstacle detection application. We show that a quite basic segmentation process applied to our disparity maps yields better results than with disparity maps computed using 2D neighbourhoods. The behaviour of the 1D method with regard to untextured areas or uncalibration defaults is studied for synthetic and real images.
Complete list of metadatas

Cited literature [102 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00543976
Contributor : Ifsttar Cadic <>
Submitted on : Tuesday, December 7, 2010 - 8:44:21 AM
Last modification on : Friday, October 23, 2020 - 4:45:31 PM
Long-term archiving on: : Saturday, December 3, 2016 - 1:11:52 AM

Identifiers

  • HAL Id : tel-00543976, version 1

Collections

Citation

Sébastien Lefebvre. Approche monodimensionnelle de la mise en correspondance stéréoscopique par corrélation - Application à la détection d'obstacles routiers. Traitement du signal et de l'image [eess.SP]. Université des Sciences et Technologie de Lille - Lille I, 2008. Français. ⟨tel-00543976⟩

Share

Metrics

Record views

242

Files downloads

1574