I have developed a hierarchical face and gaze tracking by wise-combination of Appearance-Based Trackers (ABT), which estimates predefined facial features in monocular video sequences. A non-occluded facial texture was used to estimate the eyelid position for any kind of blinking.
This hierarchy of ABT Trackers can facial motion for Eyelids, Irises, Eyebrows and Lips. The head pose is also estimated by accurate measure of the 3D spatial angles and image plane position and deep scale. I have extended the appearance model to track the iris motion, while achieving correct adaptation after saccade motion or eyelid occlusions.
On-Line Face Tracking
This method of Face Tracking does not require training neither with facial textures, shapes nor facial actions. The system learns On-Line the facial textures and illumination changes, which makes it robust to occlusions and drifting problems.Consequently, unusual faces can be also tracked either to describe 3D head pose and location or the still visual features suitable of cognitive interpretation.Interpretation of Facial Expressions
Facial expression analysis is an interesting subject due to the relevance of the expressions on human emotions. I proposed a Case-Based Reasoning (CBR) classification procedure for facial expression analysis which achieves a high recognition rate by assessing confidence for the estimated classifications.
As advantages, the training process is achieved with spontaneous facial expressions which are more natural than the forced expressions currently being used in the literature. I included the eye motion to evaluate their relevance on the facial expressions, being the first step for trust and deceit analysis. Lastly, I improved the quality of the estimations by evaluating several confidence measures.
The accurate measurement of facial actions has been also used to recognize stressful expressions in video sequences. This project has been developed for Telefonica Spain S.A. The systems extracts facial action motion by using the ABT and the stress recognition is done by using trained Single-Class Support Vector Machine.