Modélisation et suivi des déformations faciales : applications à la description des expressions du visage dans le contexte de la langue des signes
Hugo Mercier · HAL (Le Centre pour la Communication Scientifique Directe) · 2007
The face, and particularly the meaning of its expressions, plays an important role in sign languages. A few studies on facial expressions in sign language exist. This is due to the lack of description tools. In this thesis, we develop methods that allow accurate and comprehensive description of the different facial movements observed during a sign language video. We use here the Active Appearance Model formalism (AAM) in order to model face, in terms of interest point displacements and texture variations. When used with an optimization method, this formalism allows to find interest point coordinates on a face. We use here an optimization method called "inverse compositional'', that can be used to obtain accurate results in an efficient manner. In the sign language context, out-of-plane rotations and hand occlusions occur frequently. Thus, the development of robust methods is needed. It exists, for that purpose, a robust flavor of the AAM optimization methods that allow to consider the input image as being noisy. We extended it in order to detect occlusions in an automatic manner, with the assumption that the algorithm behavior in the unoccluded case is known. The algorithm result consists in 2D coordinates of each interest points in each image of a video sequence, eventually linked with a confidence value. These raw results can then be used for different applications. We thus propose to describe an expressive video sequence, at each frame, as being a linear combination of unitary facial deformations activated with different intensities. Another original application consists in a video processing that prevents the face from being identified, while keeping unchanged the meaning of its expressions.