Versatile robot vision based on features of objects: comparison of norm criterion and neural network

Recognition of shape varying objects (objects whose shapes are undetermined), termed SVOs, is an important capability that robot vision system must have in order for robots to be adoptable in realistic applications. The reason for this is apparent when one analyzes everyday scenes. One may recognize objects such as a desk, a tree, and a dog in an ordinary scene. One immediately realizes that none of these objects have fixed shapes. Even if an object has a fixed shape, its apparent shape changes with variables such as distance, orientation and shading. If a robot is to become useful in the ordinary life of human beings, it must have a vision system that is versatile enough to identify objects in such varying situations. Here, the authors report their attempt to develop such a vision system with emphasis on the identification of SVOs