Emotion Recognition Based On Multi-physiological Signals

Affective computing has received more attention in recent years.Emotion recognition is a significant component,based on various measurements such as facial expressions,speech and physiological measurement.The physiological measurement is not affected by external factors and directly controlled by the ANS system.Compared with others,it is more realistic and has received widespread attention.The current study mainly focuses on the study of unimodal analysis.However,multimodal fusion of physiological signals can complement each other.The study on multimodal fusion can make the result more robust and accurate.This study focuses on the application of peripheral physiological signals and EEG signals in emotion recognition.With selected dynamic pictures and music,four kinds of emotions,mapped into valence-arousal emotional model,were induced to establish a database.The acquired signals contained peripheral physiological signals and EEG signals.The specific analysis process included signal preprocessing,feature extraction and standardization,feature selection and feature classification.Different feature extraction methods were used to obtain the emotional characteristic,and the maximum correlation minimum redundancy feature reduction algorithm was used.The selected features were used to perform quantitative analysis of sentiment calculations by using the support vector machine.This study proposed to extract EEG features by using multivariate empirical mode decomposition,which can achieve multi-channel simultaneous analysis.The results showed that compared with time and frequency domain features,time-frequency features had better performance in emotion recognition.In this study,we used two different fusion methods to classify different kinds of emotions.For the feature fusion methods,linear fusion method and multiple kernel learning method were used.For the decision fusion methods,traditional weighted voting method and improved weighted voting method were used for fusion analysis.Compared with the unimodal recognition results,the results of these two fusion methods were improved,and the multiple kernel learning feature fusion method could have better performance.