Many robotic applications in space will require vision systems to identify objects and determine their distance and orientation. Current vision systems that identify objects by analyzing shape, texture, or motion require the power of special-purpose hardware and are subject to many difficulties as the object parameters change with distance and orientation. In this paper, these difficulties are addressed through the development of a special target label whose characteristics supply all the information needed to determine the identity, range, orientation, and geometrical characteristics of the object. The scene segmentation algorithms used in the system are tuned to recognize target labels as distinct objects. Given that the geometry of the labels is known a priori and is constant, an inverse perspective transform can then be applied to determine the range and orientation of the target label from the camera. A bar code within the label provides an index into a database describing the object to which the label is attached. The database can provide a complete description of the object, including approach and grasp locations, along with a complete CAD description of the object. The system runs on an inexpensive desk-top computer and promises to make many robot vision tasks both practical and affordable.
|