My thesis fit into the Previz project (FUI 15), which the main objective is to provide to the film maker a real-time previsualization of visual effects. In this context, it is necessary to couple the virtual (visual effects) and the real (range of players, cameras motion, ...). The aim of this thesis is to study the range of players in real-time (capture, recognition, ...). To obtain a good accuracy of capture, it is necessary to have a lot of cameras which film one or several subjects wearing markers. However, the use of markers creates problems: confined for range of players, simultaneous appearance acquisition, constrained fields of applications. Markerless methods, i.e. video based (color and depth), are divided into two categories: mono-camera and multi-camera. Mono-camera methods cause problems of ambiguities, occultation, acquisition space limited, ... To correct these problems, multi-camera methods can be used on condition that have enough cameras to cover the entire area where subjects play. Nevertheless, even if a multi-camera system is used, it is necessary to control the environment (light, sceneries, ...).
This thesis will make a contribution in the field of markerless motion capture with new algorithms in order to overcome scientific obstacles:
- almost uncontrolled environnement,
- limited number of cameras which can move,
- several subjects or occulting elements.
Potentials solutions will be studied like couple the motion capture to the movement mechanic to limit research area, use the temporality to overcome ambiguities, use the movement recognition to compensate the lack of information, ... We are committed to develop new automatic algorithms (without human intervention), respecting the real time constraint. The outcomes facilitate the motion capture during filming as well as others applications (serious game, digital entertainment, home care assitance, ...).