Reconstruction and visualization of 3D models of sports events
Pâquet, Arnaud
Promotor(s) : Van Droogenbroeck, Marc
Date of defense : 26-Jun-2019/27-Jun-2019 • Permalink : http://hdl.handle.net/2268.2/6721
Details
Title : | Reconstruction and visualization of 3D models of sports events |
Translated title : | [fr] Reconstruction et visualisation de modèles 3D d’évènements sportifs |
Author : | Pâquet, Arnaud |
Date of defense : | 26-Jun-2019/27-Jun-2019 |
Advisor(s) : | Van Droogenbroeck, Marc |
Committee's member(s) : | Embrechts, Jean-Jacques
Deliège, Adrien Barnich, Olivier |
Language : | English |
Number of pages : | 83 |
Keywords : | [en] 3D reconstruction [en] 2D pose estimation [en] multi-view 3D pose estimation |
Discipline(s) : | Engineering, computing & technology > Electrical & electronics engineering Engineering, computing & technology > Computer science |
Target public : | Researchers Professionals of domain |
Institution(s) : | Université de Liège, Liège, Belgique |
Degree: | Master : ingénieur civil électricien, à finalité spécialisée en "electronic systems and devices" |
Faculty: | Master thesis of the Faculté des Sciences appliquées |
Abstract
[en] This master’s thesis proposes a method solving multi-person 3D pose estimation with a
few calibrated camera views in an outdoor soccer environment. Despite of a significant
enhancement in recent years due to the large progress in Deep Learning, the 3D human
pose estimation in an outdoor environment still requires a lot of improvement to be
considered as resolved. Variations in lighting conditions, the movement of the camera
reducing the clearness of the image together with the occlusions between the players
on the soccer pitch and the low resolution of them make the 3D pose estimation of the
players particularly challenging. To perform accurate results, each step of the multiview 3D human pose estimator is optimized. Human detection and segmentation is
improved by eliminating irrelevant detected players via projection-based methods. An
efficient 2D pose detector suitable for estimating the skeleton of low resolution players is
used. A multi-way matching algorithm accross multiple views is introduced. Whatever
the number of players in the visible part of the pitch by the cameras, the bounding
boxes of the detected players are gathered in all camera views. The 2D skeletons of a
player are triangulated by pairs to form a 3D pose. A 3D pictorial structure (3DPS) is
applied to select the best 3D pose combination. In the case of inaccurate 3D skeleton,
a bundle adjustement is performed to refine the 3D pose. The proposed approach gives
accurate results in an outdoor environment.
File(s)
Document(s)
Description: -
Size: 10.69 MB
Format: Adobe PDF
Cite this master thesis
The University of Liège does not guarantee the scientific quality of these students' works or the accuracy of all the information they contain.