Are You Talking to Me? Detecting Attention in First-Person Interactions
Luis Carlos González-García, Luz Abril Torres-Méndez, Julieta Martínez, Junaed Sattar, James J. Little · 2015
This paper presents an approach for a mobile robot to detect the level of attention of a human in first-person interactions. Determining the degree of attention is an essential task in day-to-day interactions. In particular, we are interested in natural Human-Robot Interactions (HRI’s) during which a robot needs to estimate the focus and the degree of the user’s attention to determine the most appropriate moment to initiate, continue and terminate an interaction. Our approach is novel in that it uses a linear regression technique to classify raw depthimage data according to three levels of user attention on the robot (null, partial and total). This is achieved by measuring the linear independence of the input range data with respect to a dataset of user poses. We overcome the problem of time overhead that a large database can add to real-time Linear Regression Classification (LRC) methods by including only the feature vectors with the most relevant information. We demonstrate the approach by presenting experimental data from humaninteraction studies with a PR2 robot. Results demonstrate our attention classifier to be accurate and robust in detecting the attention levels of human participants. Keywords–Human-robot interaction; Body pose classification; Least squares approximations; Raw range data analysis.