A method of multi-factorization for recognizing emotions from gestures

We propose a new method of recognizing emotional factors from human gestures by analyzing motion capture (MoCap) data. It features multi-factorization processing combined with HMM recognition. The multi-factorization processing factorizes MoCap data into a third-order tensor that consists of spatial, statistical, and frequency-spatial components. This multi-factorization localizes the data in the factorized tensor space according to their mutual correlation, which results in helping data clustering. This means that the proposed tensor-shaped features have advantages over conventional features in recognizing emotions from gestures. The validity of the proposed method was confirmed using the results of experiments in which emotions from walking actions were analyzed.