Reconnaissance des expressions faciales à base d’informations vidéo ; Estimation de l’intensité des expressions faciales

Abstract

Facial expressions recognition is a basic task in the human machine communication. During this thesis, we developed a facial expression classification system based mainly on transient features. In order to carry out this system, we considered two previous works developed in the laboratory. These two works relate to the detection of facial permanent features which are eyes, eyebrows and lips. This automatic detection allowed the localization of characteristic points of these features which are the corners and the limits of each feature. From these characteristic points, we detected facial regions where transient features can appear. A study on the presence or the absence of this type of features s well as other characteristics of this type of features gave a primary classification of facial expressions without providing much effort. This study is completed by a study on permanent features in order to refine obtained results. This categorical classification proved its limit in the recognition of other facial expressions, this leads to develop another system which propose a dimensional classification of facial expressions in terms of positivity / negativity or pleasure / not pleasure. An ideal facial expression system allows quantifying facial expressions. Another aim of this thesis is the development of a quantification system which allows estimating intensity of known as well as unknown facial expression. The three suggested systems are based in addition to the transient features, on different deformations of permanent features. To measure the deformation of these features, five characteristic distances are computed from characteristic points located from detected permanent features. The Transferable Belief Model is used in the three systems in order to fuse all available data coming from different sensors. This is why we have used this specified model. Another raison of the choice of this model is its ability to model the doubt between the different considered classes. Another objective of this thesis is the recognition of facial expressions based on video information. When considering facial expressions databases, several actors present different expressions with various intensities in video sequences. Each video count more than 200 frames. Several data are extracted from each frame this leads to an important mass of data. To analyze and explore these data we use a data mining technique in order to extract a new temporal knowledge. The developed dynamic recognition system takes into the count the temporal deformations of face features. It induces temporal rules which describe facial expressions in a dynamic context. These rules can be added to existing ones proposed by M-PEG4 in a static context.