Low cost human-robot gaze estimation system

摘要

This work presents the development of a low cost Human-Robot gaze estimation system for the purpose of promoting joint Human-Robot workspaces in daily scenarios. We have developed this system using only monocular eye tracking and the 2D gaze point. Prior to this work there have been many efforts to bridge the gap between human and robots by developing robotic assistants capable of serving humans through command based interaction or manual input. Such systems lack freedom and require constant input whenever interaction is required. The system we have developed allows for both seated and free roaming interaction between a human and the robot. The end software allows the robot to track the gaze point of the human in real-time and fixate on the same point in space. Once the robot is stabilized on this gaze point it may perform a number of different tasks which may assist the human. We have constructed our own pair of gaze tracking glasses using only low cost components to demonstrate effective performance under budget costs. These glasses are coupled with a pan/tilt laser robot which tracks the gaze point of the human on the environment by projecting a red laser. We have designed 2 separate subsystems which track the pose o the user's head in real time and then combine to give an accurate estimate relative to the robot. As well as this we have developed a gaze tracking module which calibrates the gaze point of the human to a relative gaze window on the robot. Data from all 3 subsystems are then combined through a data fusion model and then sent to the robot to adjust its pan/tilt angles to focus on the same point in space as the human. These subsystems combine to provide an accuracy of 94% to the centre of a target object. The system was tested through a user study which involved 12 subjects undergoing 5 different testing scenarios.