Abstract
This paper describes a new approach to visual servoing. The visual servoing system has a camera mounted on a manipulator, which is included in a control loop of the manipulator as a visual sensor observing the objects. The key problem in visual servoing is a reduction of dimensionality of image data without a loss of sensitivity to change of object appearance. We use eigenspace method based on Karhunen-Loève expansion which reduces the dimensionality and characterizes a given image data set. The proposed method does not require feature extraction, feature correspondences or object models for obtained images. It relates the camera pose and the reduced image data directly, and leads the robot arm mounting the camera to a goal position.
We construct a surface (manifold) which constraints the camera position and the dimension reduced image in their product space. Using tangential properties, we control the camera to a given goal position which is specified as where the goal image is obtained. Experiments using a real robot arm and a TV camera show feasibility of this proposed method.
