Reconstruction 3D de scènes d'intérieurs à partir de photographies

Abstract : The 3D reconstruction of many objects and/or scenes from their photographies is made possible by current photogrammetry methods. To do so, usual methods detect salient points in every pictures and then match them between each pictures. These matches then give information on the position of every camera that took a picture of the scene. Once these positions are obtained, a dense reconstruction of the scene can be obtained by triangulating the parts seen in different pictures. Point detection and matching are crucial parts of these 3D reconstruction methods. That is why some scenes or objects are still hard to reconstruct in 3D with photogrammetry methods. Indoor scenes belong to these difficult cases, with their lack of texture that causes point detection and matching to give poor results. Moreover, the planarity of these scenes is a degenerate case for usual calibration methods. Combined, these drawbacks explain the difficulty to calibrate such scenes. In this thesis, we explain how to use segments to compensate for the lack of robustness of point methods in the case of indoor scenes. First, we introduce a segment detection method that is more robust to the lack of contrast in indoor scenes. This multi-scale method also gives good results whatever the image resolution is. We use criterion inspired from emph{a contrario} methods to avoid usage of several parameters. We then present a bifocal calibration method that uses both line segments and points. Segments allow the method to still work in low-texture and/or planar scenes and points allow the method to be as accurate as other point methods. To do so, we introduce an emph{a contrario} RANSAC variant to choose, for each scene, whether points or line segments should be used for calibration. Finally, to deal with the lack of overlap between consecutive pictures in indoor scenes, we introduce a multi-view calibration method that uses coplanarity constraints between segments when there are no trifocal constraints. We explain how to modify usual trifocal constraints to combine them with coplanarity constrains in order to obtain a method as accurate as usual methods but more robust in wide-baseline scenes
Document type :
Theses
Complete list of metadatas

Cited literature [80 references]  Display  Hide  Download

https://pastel.archives-ouvertes.fr/tel-01769090
Contributor : Abes Star <>
Submitted on : Tuesday, April 17, 2018 - 4:49:16 PM
Last modification on : Thursday, May 16, 2019 - 6:07:03 PM

File

TH2017PESC1186_diffusion.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01769090, version 1

Citation

Yohann Salaün. Reconstruction 3D de scènes d'intérieurs à partir de photographies. Traitement des images [eess.IV]. Université Paris-Est, 2017. Français. ⟨NNT : 2017PESC1186⟩. ⟨tel-01769090⟩

Share

Metrics

Record views

158

Files downloads

75