Based On User-independent Model For Emotion Recognition Using Physiological Signals

Emotion recognition technology using physiological signals has been applied ineducation, medical care and other aspects and the system based on users-dependentmodel achieved satisfactory results. But the performance of user-independent systemthat the practical applications required was not good enough. When added by a newperson’s test data, the system classifying rate decreases significantly. So, building auser-independent model and choosing common features with good classificationability are pervasive problems to be solved.Aiming to solve the problem referred above, several volunteers’ ECG and GSRwere collected, during the experience of four kinds of basic emotions induced byvideos. After preprocessing, features were extracted, and build up biological featuredatabase. Features were normalized using std., and reduced using PCA. According tothe distribution of features in principal component space and the distance of sampleswas calculated. Experiment on5subjects’ data results that the distances ofinter-classes have big different from within-classes. The recognition accuracy undercross-validation reaches up to0.99, yet, less than50%for leaving one subject outvalidation. The results indicate the advance of user-dependent model in full samplingspace, but not available in small sampling rate systems.This article uses the correlation and spectral clustering to analysis featureredundancy, and uses Pairwise constraints score to grade the classification ability offeatures. The results of experiment on Aubt database and one person’s data and all the12subjects’ data show that, features of P, Q, R, S, T extract form ECG are fullycorrelated, so are the features of GSR. ECG periodic features have strongclassification ability, while GSR features are weak, which were proved by Aubt anddatabase with one person. However, the grades of each feature are different betweenthe two databases. Results of12subject’s database show all features have nosignificant difference. The results illustrate that features extracted by this experimentis not universal, because of individual personality physiological space.User-independent model proposed in this paper will perform excellent in userfixed recognition system; however, the small sampling rate system’s performance willbe not satisfying. This problem can be improved by increasing the coverage of thesample, and improving the personality parameters. Features are highly redundant andlittle universal. Feature work should appropriately selected features to reduce featureredundancy, do physiological baseline calibration, and transform individual space to auniversal space according to the individual physical, psychological information.