ACTIVE 3D VISION THROUGH GAZE RELOCATION IN A HUMANOID ROBOT

Abstract

Motion parallax, the relative motion of 3D space at different distances experienced by a moving agent, is one of the most informative visual cues of depth and distance. While motion parallax is typically investigated during navigation, it also occurs in most robotic head/eye systems during rotations of the cameras. In these systems, as in the eyes of many species, the optical nodal points do not lie on the axes of rotation. Thus, a camera rotation shifts an object's projection on the sensor by an amount that depends not only on the rotation amplitude, but also on the distance of the object with respect to the camera. Several species rely on this cue to estimate distance. An oculomotor parallax is present also in the human eye, and during normal eye movements, displaces the stimulus on the retina by an amount that is well within the range of sensitivity of the visual system. We developed an anthropomorphic robot equipped with an oculomotor system specifically designed to reproduce the images impinging on the human retina. In this study, we thoroughly characterize the oculomotor parallax emerging while replicating human eye movements and describe a method for combining 3D information resulting from pan and tilt rotations of the cameras. We show that emulation of the dynamic strategy by which humans scan a visual scene gives accurate estimation of distance within the space surrounding the robot.