Construction and cross-correlation analysis of the affective physiological response database

Six kinds of emotions, namely joy, surprise, disgust, grief, anger and fear, are elicited from 300 ordinary college students by means of high-rated movie clips. The signals of galvanic skin response (GSR), heart rate (HR), blood volume pulse (BVP), electrocardiogram (ECG), respiration (Rsp.), facial electromyography (EMG) and two channels of electroencephalography (EEG) from the frontal lobe were recorded while the subjects were watching the affective eliciting materials. The samples of each kind of signals are analyzed by using random matrix theory (RMT). From the distribution of the eigenvalues and eigenvectors of the correlation matrices of BVP, facial EMG and two EEGs, we find that the statistical properties of these four kinds of signals coincide with the RMT predictions. However, the largest eigenvalue and the corresponding eigenvector component distribution of the correlation matrices of GSR, HR, ECG and Rsp. have deviated significantly from the RMT predictions. Our experiments also reveal that the correlations between two arbitrary samples of BVP, facial EMG or EEGs does not result from affective response but from random noises, whereas the amplitude variations of GSR, HR, ECG and Rsp. contain the correlated affective physiological response patterns. According to the data analysis results of RMT, 110 features are firstly extracted from the GSR and HR signals, and then a subset of the initial features is selected by using the Backward Selection algorithm. The affective data samples each of which is a vector of the features are classified by using a Fisher classifier, and furthermore, the user-independent emotion recognition systems with good prediction performance are constructed during the model selection process.