Abstract
abstract: In this paper, application of an eigenspace method to a visual manipulation system is described. In contrast to traditional visual control approaches, visual information captured by a CCD camera is directly processed in the proposed scheme, using the so called eigenspace method rather than shape recognition. A set of images of the controlled object obtained by being coarsely sampled from the workspace is compressed to a low-dimension space called as an eigenspace. Thus every image of the controlled object can be mapped to a point in the eigenspace. So, in another meaning, every state of the object corresponds to a point in the eigenspace. Moreover, a new method using a Radial Basis Functions Network to project the image's eigenspace coordination to object's state is developed. Thereby using a suitable eigenspace calculated in advance, a new image captured by the camera can be mapped into the eigenspace, and the corresponding object's state can be estimated during operation. As examples, a real-time visual manipulation system to perform the so-called “ball and beam” control task is adopted. Finally, performance of the scheme is examined by experiments.
