Effects of Emotion Grouping for Recognition in Human-Robot Interactions

Abstract

Understanding people's emotions may be important to achieve success in behavior adaptability and, consequently, to sustain long-term human-robot interactions. Most emotion recognition systems consist in classifying a given input into one out of seven basic emotions, following Ekman's model. However, it is sometimes enough for the customization of a robot's behavior to recognize whether an emotion is positive or negative, in order to approach more often subjects which display more positive emotional reactions. In this article, two approaches to that effect are proposed and compared. The first one, named pre-grouping, refers to combining the four negative emotions into one single class and use it to train a classifier. The second one, named post-grouping, refers to applying classifiers to classify the seven basic emotions and interpret their negative outputs as related to a single class. Furthermore, a novel dataset entitled QIDER, based on queries in a search engine and well defined facial cues, is introduced and made available for public use. Both approaches led to more balanced precision scores among all classes, which may make them a suitable choice for applications in human-robot interaction. Several experiments have been performed and post-grouping is shown to produce better overall accuracy.