Estimating target orientation with a single camera for use in a human-following robot

This paper presents a monocular vision based technique for extracting orientation information from a human torso for use in a robotic human-follower. Typical approaches to human following use an estimate of only human position for navigation, but we argue that a better navigation scheme should include directional information. We propose that the pose of a walking person’s upper body typically indicates their intended travelling direction, and show that a simple planar fit to the back of a human torso contains sufficient information for the purpose of inferring orientation. We obtain this planar fit using only 2D image points. Results showing the efficacy of this approach are presented, together with those of a simple human-following controller incorporating the pose estimate.