Interaction Intention Recognition via Human Emotion for Human-Robot Natural Interaction

In many social scenarios, human emotion governs external behavior, and behavior reflects intention. If a social robot can recognize users' interaction intention through observable behaviors, it can respond in a personalized way, thus exhibiting "natural" behavior. In this paper, we propose a new method for social human-robot interaction, where the robot obtains the facial expression and body action of the people only from RGB videos to recognize whether humans intend to interact with it or not. Through feature extraction and fusion, a computational model of emotion is established to further identify the interlocutors' intention. We use a stacking model with 3 basic classifiers to fuse features and create a real unconstrained dataset to evaluate our method. The result shows that our method which combines facial expression and body action features, is more efficient than single-cue based methods with respect to increased accuracy, robustness, and reduced uncertainty.