Skip to Main content Skip to Navigation

Mise en correspondance stéréoscopique d'images couleur en présence d'occultations

Sylvie Chambon 1, 2 
Abstract : This work deals with stereo-vision and more precisely matching of pixels using correlation measures. Matching is an important task in computer vision, the accuracy of the three-dimensional reconstruction depending on the accuracy of the matching. The problems of matching are: intensity distortions, noises, untextured areas, foreshortening and occlusions. Our research concerns matching color images and takes into account the problem of occlusions.

First, we distinguish the different elements that can compose a matching algorithm. This description allows us to introduce a classification of matching methods into four families : local methods, global methods, mixed methods and multi-pass methods.

Second, we set up an evaluation and comparison protocol based on fourteen image pairs, five evaluation areas and ten criteria. This protocol also provides disparity, ambiguity, inaccuracy and correct disparity maps. This protocol enables us to study the behavior of the methods we proposed.

Third, forty correlation measures are classified into five families : cross-correlation-based measures, classical statistics-based measures, derivative-based measures, non-parametric measures and robust measures. We also propose six new measures based on robust statistics. The results show us the most robust measures near occlusions : the robust measures including the six new measures.

Fourth, we propose to generalize dense correlation-based matching to color by choosing a color system and by generalizing the correlation measures to color. Ten color systems have been evaluated and three different methods have been compared : to compute the correlation with each color component and then to merge the results; to process a principal component analysis and then to compute the correlation with the first principal component; to compute the correlation directly with colors. We can conclude that the fusion method is the best.

Finally, in order to take into account the problem of occlusions, we present new algorithms that use two correlation measures: a classic measure in non-occluded area and a robust measure in the whole occlusion area. We introduce four different methods: edge detection methods, weighted correlation methods, post-detection methods and fusion method. This latter method is the most efficient.
Document type :
Complete list of metadata

Cited literature [236 references]  Display  Hide  Download
Contributor : Sylvie Chambon Connect in order to contact the contributor
Submitted on : Friday, January 20, 2006 - 5:34:13 PM
Last modification on : Friday, August 5, 2022 - 2:56:21 PM
Long-term archiving on: : Saturday, April 3, 2010 - 9:37:59 PM


  • HAL Id : tel-00011437, version 1


Sylvie Chambon. Mise en correspondance stéréoscopique d'images couleur en présence d'occultations. Interface homme-machine [cs.HC]. Université Paul Sabatier - Toulouse III, 2005. Français. ⟨tel-00011437⟩



Record views


Files downloads