A GMM based 2-stage architecture for multi-subject emotion recognition using physiological responses

There is a trend these days to add emotional characteristics as new features into human-computer interaction to equip machines with more intelligence when communicating with humans. Besides traditional audio-visual techniques, physiological signals provide a promising alternative for automatic emotion recognition. Ever since Dr. Picard and colleagues brought forward the initial concept of physiological signals based emotion recognition, various studies have been reported following the same system structure. In this paper, we implemented a novel 2-stage architecture of the emotion recognition system in order to improve the performance when dealing with multi-subject context. This type of system is more realistic practical implementation. Instead of directly classifying data from all the mixed subjects, one step was added ahead to transform a traditional subject-independent case into several subject-dependent cases by classifying new coming sample into each existing subject model using Gaussian Mixture Model (GMM). For simultaneous classification on four affective states, the correct classification ration (CCR) shows significant improvement from 80.7% to over 90% which supports the feasibility of the system.