Abstract
Useful information about affects and affective states can be extracted form the physiological signals, even under difficult lighting and pose conditions. Little work has been done on using physiological signals in automated affect recognition systems. We employed measurements of facial skin temperature variations for developing a non-intrusive automated facial expression classification system. Variances in thermal intensity values recorded at thermally significant locations on human faces were used to discern between normal, pretended happy and pretended sad facial expression of affective states. A three-step algorithmic approach was used to construct the classifier. Employed approach involved derivation of principal components, stepwise selection of optimal and most discriminating features, and linear discriminant analysis within the reduced optimal eigenspace. The resulting classifier performed at an impressively low (16.2%) error rate