This manuscript deals with the problem of markerless motion capture. An approach to thisproblem is model-based and is divided into two steps : an initialization step in which the initialpose is estimated, and a tracking which computes the current pose of the subject using infor-mation of previous ones. Classically, the initialization step is done manually, for bidding the possibility to be used online, or requires constraining actions of the subject. We propose an automatic real-time markerless initialization step, that relies on topological information provided by skeletonization of a 3D reconstruction of the subject. This topological information is then represented as a tree, which is matched with another tree used as modeldescription, in order to identify the different parts of the subject. In order to provide such a method, we propose some contributions in both digital topology and graph theory researchfields. As our method requires real-time computation, we first focus on the speed optimization of skeletonization methods, and on the design of new fast skeletonization schemes providing good results. In order to efficiently match the tree representing the topological information with the tree describing the model, we propose new matching definitions and associated algorithms. Finally, we study how to improve the robustness of our method by the use of innovative con-straints in the model. This manuscript ends by a study of the application of our method on several data sets, demon-strating its interesting properties : fast computation, robustness, and adaptability to any kindof subjects
| Identifer | oai:union.ndltd.org:CCSD/oai:tel.archives-ouvertes.fr:tel-00597513 |
| Date | 07 December 2010 |
| Creators | Raynal, Benjamin |
| Publisher | Université Paris-Est |
| Source Sets | CCSD theses-EN-ligne, France |
| Language | English |
| Detected Language | English |
| Type | PhD thesis |
Page generated in 0.0028 seconds