Emotion recognition using EEG signals

• Main Author: Soh, Wei Lin
• Other Authors: Smitha Kavallur Pisharath Gopi
• Format: Final Year Project
• Language: English
• Published: 2015
• Subjects: DRNTU::Engineering::Computer science and engineering::Computer applications::Social and behavioral sciences
• Online Access: http://hdl.handle.net/10356/65603
• Institution: Nanyang Technological University

Emotions are essential in non-verbal communication between people and yet, they are complex and hard to interpret accurately. Electroencephalogram (EEG) signals, being physiological signals and have faster data collection process, could give a better indication of the emotions one is experiencing. Despite various researches done on emotion recognition using EEG signals, those that used music videos as stimulus attained lower accuracy of 58.8% and 55.7% for valence and arousal dimension respectively. However, music videos, being both audio and visual stimulus, are more similar to stimulus that elicits emotions in people in everyday’s life. Therefore, this study aims to design and implement an automatic emotion recognition system using EEG signals with music videos as stimulus. Bandpower and bandpower asymmetry are often extracted as features in the alpha and beta band to determine arousal and valence respectively, but it was not always the same electrodes in the researches that give good classification accuracy. As such, this study also analyses EEG signals in the frequency domain to determine the effective electrodes to recognize emotions in the music videos context. EEG signals were collected using Emotiv Epoc, application to elicit emotions was written using C# .NET and processing of data was done in Matlab. Using Support Vector Machine (SVM), results showed that bandpower extracted from electrodes AF3, F3 and AF4 in the alpha band and electrodes T7, P8 and T8 in the beta band and bandpower asymmetry extracted from (AF3-AF4) and (F3-F4) in the alpha band and (T7-T8) and (P7-P8) in the beta band gave us the best average classification accuracy of 73.3%. This was an improvement from 62.8%, when temporal window length is 0-10 seconds, instead of 2-10 seconds. This suggests the possibility of taking into account individuals’ response differences to the stimulus or improving the database of music videos stimulus to further improve on the classification accuracy.