Human Emotion Recognition Using Physiological Signals

Abstract

Understanding the emotional state of an individual has become the latest attraction of research in the field of human-computer interaction systems. By recognizing the emotional state of the user, these systems will help to take more complex and efficient decisions. The use-cases include providing assistance to individuals with severe motor disabilities, predicting automobile drivers’ subsequent actions, making better e-learning systems and the list goes on. This research work is mainly based on designing efficient models for the task of human emotion recognition utilizing the responses of their physiological signals to some external stimuli. Several models of the two classifiers namely, k-Nearest Neighbour(k-NN) and the Hidden Markov Model(HMM), have been proposed and implemented in the current work. A subset of the DEAP dataset, considering only the EEG, EMG, GSR, RSP, TEMP physiological signals, is used for training the models, considering the classification as a two-class problem viz. low-arousal and high-arousal. Discrete Wavelet Transform(DWT) has been used both for signal pre-processing as well as for signal feature extraction. Also, the relative wavelet sub-band energy is used as features for two of the implemented models. A comparative analysis of the models, considering only EEG signals and EEG signals along with the peripheral signals, has been done. It is found that out of the three implemented models, the HMM classifier has a pretty good accuracy of about 72.5% for subject-independent classification. Besides the detailed explanation and discussions about the implementation of the proposed classifier models, this thesis also presents a consolidated study of the work being carried out in this comparatively new area of study of emotion recognition.