Quantitative intensity analysis of facial expressions using HMM and linear regression

Abstract

In this paper, an automatic framework of facial expression analysis focusing on quantitative intensity illustration is proposed. Quantitative intensity variation could be extracted during the whole period of facial expressions, from neutral state to apex state. The logic behind this paper lies in the intensity differences of same prototype expression, and lies that these intensity differences could be illustrated by facial expression energy variation throughout expression. In order to unify video data with different frame numbers, Hidden Markov Models (HMMs) are applied to every video for classification and expression states generation. These expressions states extracted from each video showing same expression have the same length. Then given facial landmarks of key positions, energy value of each state could be demonstrated by placements of landmarks. By synthesizing states variation and energy value, intensity curves for each expression could be obtained using linear regression algorithm. In this work, we explore person-dependent and person-independent analysis of expressions, in person-dependent experiment quantitative intensity compare is tested for expression 'Happiness'.