Combining the Frontal and Side AAM For Robust Face Detection

The statistical modelling of faces using Active Appearance Models is an efficient approach to detect and interpret faces. Two important drawbacks of the method are the lack of robustness against occlusions and side poses of the face. The FR-PCA algorithm introduced recently provides a successful method of getting rid of occlusions in the image by reconstruction it using PCA bases of random samplings. In the context of this thesis, the Side AAM for modelling profile faces were introduced, and the FR-PCA algorithms usage was extended by employing some specific features of the reconstruction to build a classifier that can detect occlusions and to design a second method that can decide on the pose of the face in a similar manner. Finally these three approaches were combined together and a robust model of face detection was built, implemented and verified. The model is robust against occlusions and it can switch between frontal and side AAMs with the help of the occlusion and pose detectors proposed.