A Feature Learning Approach Based on Multimodal Human Body Data for Emotion Recognition

The human body as an entire structure of a person contains physiological and physical reactions that connect with emotions. Emotions play a crucial role in our day to day activities, not only the way we interact with colleagues but also in our decision making processes. Emotion recognition from multimodal signals allows a direct assessment of the innermost state of a person which is regarded an important component of Human-Computer interactions (HCI). This paper proposes an enhancement feature learning technique using a Hyper Enhanced Learning System (HELS). We conduct an experiment on the DEAP database by utilizing four signals. Signals were preprocessed to remove artifacts and noise and feature extraction is done to obtain relevant features. We then introduced a feature enhancement system that generates random weights and enhances feature nodes. We assorted emotions into three categories and classified the emotional states using Artificial Neural Networks (ANN). The average accuracy rate of our proposed system was 78.6% and 79.9% respectively for valence and arousal. Obtained results show that combining physiological signals is relevant for accurate human emotion recognition task.