A Novel Cross-FOV Gaze-Driven Human-Robot Interaction Framework for Service Robots.

Gaze-based methods have been widely applied in human-robot interactions (HRI). However, most methods were often limited to scenarios where the human face and the gazed object should be in the same field of view (FOV), which is difficult to adapt to the complex and changeable typical daily life scene where the service robot is located. To improve the limitation, this paper proposes a framework capable of completing tasks even if the person and the gazed object are not in the same FOV, requiring only a single RGB camera. First, a cross-FOV gaze estimation algorithm is designed to get the gaze point from images containing a human face captured by an RGB camera. Then a multi-view fusion decision model is employed to drive the robot to turn to the gaze point and identify the object according to the gaze point. Finally, a robot can complete tasks using this information. Experiments under conditions of different lighting, different scenes and different users on the NAO robot show the effectiveness of our method.