Emotional State Recognition Based on Physiological Signals
Name
Artem Bachynskyi
Abstract
Emotional state recognition is a crucial task for achieving a new level of Human-Computer Interaction (HCI). Machine Learning applications penetrate more and more spheres of everyday life. Recent studies are showing promising results in analyzing physiological signals (EEG, ECG, GSR) using Machine Learning for accessing emotional state. Commonly, specific emotion is invoked by playing affective videos or sounds. However, there is no canonical way for emotional state interpretation. In this study, we classified affective physiological signals with labels obtained from two emotional state estimation approaches using machine learning algorithms and heuristic formulas. Comparison of the method has shown that the highest accuracy was achieved using Random Forest classifier on spectral features from the EEG records, a combination of features for the peripheral physiological signal also shown relatively high classification performance. However, heuristic formulas and novel approach for ECG signal classification using recurrent neural network ultimately failed. Data was taken from the MAHNOB-HCI dataset which is a multimodal database collected on 27 subjects by showing 20 emotional movie fragment`s. We obtained an unexpected result, that description of emotional states using discrete Eckman's paradigm provides better classification results comparing to the contemporary dimensional model which represents emotions by matching them onto the Cartesian plane with valence and arousal axis. Our study shows the importance of label selection in emotion recognition task. Moreover, obtained dataset have to be suitable for Machine Learning algorithms. Acquired results may help to select proper physiological signals and emotional labels for further dataset creation and post-processing.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Ilya Kuzovkin
Defence year
2018