Subject-Dependent and Subject-Independent Emotion Classification Using Unimodal and Multimodal Physiological Signals

摘要

Our work is aimed to investigate the feasibility and suitability of subject-dependent and subject-independent emotion classification using unimodal and multimodal physiological signals. We propose to use EEG, ECG, and SC to classify valence and arousal emotion using SVM classifier. The emotions are elicited by pictures and classical music. In classifying valence, the best subject-dependent and subject-independent accuracy are 84.44% and 75.00% respectively. In classifying arousal, the best subject-dependent and subject-independent accuracy are 79.44% and 69.44% respectively. Considering in almost every aspect, most of subject-independent accuracies are worse than subject-dependent accuracies, probably because of inter-participants variability. The most accurate unimodality in classifying valence and arousal is EEG and ECG respectively. The most accurate multimodality in classifying valence and arousal is feature-level fusion and decision-level fusion respectively. The most accurate modality in classifying valence and arousal is EEG and decision-level fusion respectively. Considering the number of significant features, selected by ANOVA, EEG and SC have tendency to classify valence better than arousal while ECG has tendency to classify arousal better than valence.