Abstract
We propose a novel feature extraction and pattern representation approach to develop classification rules for affect recognition using bio-physiological data from a training sample of individuals asked to express various emotions by natural facial gesture. Thermal imaging features at corresponding facial locations were shown to follow a multivariate normal distribution, with clustering at different centres in feature space for different positive or negative affective states. A multivariate analysis of variance was used to represent thermal images as vectors with components along principal components (PCs) of a covariance matrix. Derived PCs were ranked in order of their effectiveness in the between-cluster separation of affective states, and only the most effective PCs were retained in an optimized subspace for affect recognition purposes. A set of Mahalanobis distance based rules was constructed to classify the simulated affective states in a person-independent manner. Results suggest that an optimized subspace allows better between-cluster separation, and hence better emotion detection by, for example, a robot, than standard reduction to PCs contributing most of the training sample variance.