Syntactic pattern recognition for robot vision

Vision will be increasingly important in robotics applications. An object is modeled by critical curves extracted from the object. In the simplest case the object's curve is its outline. A solution to the curve partitioning problem is shown for nonconvex objects. Curves are described in a rotation and translation invariant way. A way to build a combined model database of many classes of objects is presented. A test to insure disjointness of model classes is given. An efficient technique for computing a network of these curves from a gray-level frame is presented. A graph algorithm is presented to match the model in an efficient way, independent of the scaling found in the scene. A technique for computing and classifying more general critical curves from three-dimensional data is developed.