BASIS OF EIGENFACES FOR TRACKING OF HUMAN HEAD

Ján Mihalík - Miroslav Kasár

   The paper deals with the construction of the basis of eigenfaces from a training set of face images using the wireframe 3D model Candide of the human head. The wireframe 3D model is adapted to the face images from the training set, because in general they are obtained from different sources. Afterwards the face images with the adapted model are geometrically normalized to remove texture variations caused by their global and local motion and geometrical differences between individuals. As a result there are more similar geometrically normalized textures of the human faces than before the normalization process. To get completely pre-processed textures of the human faces after their geometrical normalization they are centred and energetically normalized to minimize the effect of global lighting variation. By applying the principal component analysis to these completely pre-processed textures the basis of eigenfaces is obtained. Finally the basis of eigenfaces combined with an algorithm of adapting of 3D model is applied on tracking of the human head in head-and-shoulder videosequence.

Keywords: eigenfaces, 3D model, Candide, texture, shape, principal component analysis, tracking system

[full-paper]