Skip to Main content Skip to Navigation

Positionnement visuel dans un monde d'objets

Vincent Gaudillière 1, 2 
1 MAGRIT-POST - Augmentation visuelle d'environnements complexes
Inria Nancy - Grand Est, LORIA - ALGO - Department of Algorithms, Computation, Image and Geometry
2 MAGRIT - Visual Augmentation of Complex Environments
Inria Nancy - Grand Est, LORIA - ALGO - Department of Algorithms, Computation, Image and Geometry
Abstract : Augmented Reality can be defined as the superimposition of reality and elements (sounds, 2D and 3D images, videos, etc.) calculated in real time by a computer system. In practice, this term refers to the addition of visual elements, either in the field of view of an observer through specific glasses (e.g. Microsoft Hololens, Magic Leap One), or on a screen through which the observer sees reality (usually a smartphone or tablet). During this research work, we were interested in the deployment of Augmented Reality in an industrial context, and more particularly in the challenges that large industrial environments (factories, plants, ships) represent in terms of image analysis and processing. In particular, we investigated the use of objects of interest present in the scene to recognize the place of the observer and then calculate his precise position with respect to the environment. Applications include manufacturing assistance, maintenance assistance, documentation and training. After proposing a functional definition of the concept of place in an industrial environment, as a zone of interaction around an object of interest, we approached place recognition as an image retrieval task in which the similarity between the unknown image and the reference images is measured in two steps. The validity of the images with the greatest similarity to the unknown image is then assessed by epipolar geometry estimation between the unknown image and each of the retrieved images. The similarity measurement and geometry estimation are guided by the calculation of object-level correspondences between regions of interest of the two images. To calculate the camera pose, we then took advantage of the objects of interest present in the scene, using a modeling of the latter in the form of ellipsoids, the projections of the objects in the image being modeled as ellipses. Our contributions to the problem of estimating camera pose from ellipse - ellipsoid correspondences are both theoretical and practical. In particular, we have shown that there is a parametrization of the solutions to the one-ellipsoid problem, and, moreover, that the camera pose estimation problem can be reduced to an orientation estimation problem only. We have also proposed a robust way to handle the multiple possible matches between the objects detected in the image and the objects present in the 3D scene model.
Complete list of metadata

Cited literature [104 references]  Display  Hide  Download
Contributor : Vincent Gaudilliere Connect in order to contact the contributor
Submitted on : Friday, August 21, 2020 - 4:21:38 PM
Last modification on : Saturday, June 25, 2022 - 7:43:04 PM


Files produced by the author(s)


  • HAL Id : tel-02915866, version 4


Vincent Gaudillière. Positionnement visuel dans un monde d'objets. Vision par ordinateur et reconnaissance de formes [cs.CV]. Université de Lorraine, 2020. Français. ⟨NNT : 2020LORR0063⟩. ⟨tel-02915866v4⟩



Record views


Files downloads