Head Pose Estimation is an Inadequate Replacement for Eye Gaze in Child-Robot Interaction

Abstract

Gaze analysis of human-robot interactions can reveal much about the dynamics of the interaction and be a useful step in establishing levels of engagement and attention. Currently, much of this work has to be conducted manually through post-hoc video coding due to current limitations in non-invasive, real-time gaze tracking solutions. This paper assesses whether real-time head pose estimation from an RGB-D camera may be used in place of manual post-hoc coding of gaze direction. Using data collected from an experiment `in the wild', it is found that the proposed RGB-D based pose estimation method is neither accurate nor consistent enough to provide a reliable measure of gaze within human-robot interactions.