Calibrating the Eye Motion of an Humanoid Robot

In the human visual system, the projective relationship between the images seen in each eye with each other changes with their motion as the viewer attends to different points in space. The active vision heads built for many humanoid robots approximate human gaze behavior and share this property. Knowledge of this projective relationship is used in stereo vision tasks and is captured entirely by the position and orientation of the cameras with respect to each other and properties intrinsic to the individual cameras. In this work we present a method for inferring a kinematic model of a robots active vision system and use it to estimate of the system's epipolar geometry as it changes when the cameras undergo motion. This kinematic model constitutes a model of the self in terms the visual system. Epipolar geometry describes the projective relationship between the two cameras in a stereo vision system. We model these cameras using the pinhole camera model. In a classical pinhole camera, we are able to take pictures because a pinhole in the center of the camera allows only a narrow ray of light to pass through, carrying the color information of the object that it bounced off of prior to entering the camera. Therefore, where a single pixel imaged by a pinhole camera resides in 3-space is constrained to a single ray of light. In a stereo pair of cameras, this ray of light will lie along a single line in the other image. This is referred to as the epipolar constraint, so named because all such epipolar lines pass through the epipole, the image of the other camera's camera center[1]. Estimating this relationship is referred to simply as the estimation of epipolar geometry, whereas estimating those parameters intrinsic to the camera is referred to as camera calibration. An active vision system is one in which either the cameras are able to move or in which they include manipulators that are able to interact with the environment. An example of such a vision system is a camera attached to a movable arm, referred to as a hand camera. Hand-Eye and Head-Eye calibration both refer to the process of inferring the mounting of a camera to an underlying system with known kinematics. If these kinematics are unknown, the process of inferring them is kinematic calibration. Nico is an upper-torso humanoid robot modeled after the fiftieth percentile 12-month-old male infant. The …