System for projectively invariant recognition of planar objects
Peter Vanroose, Luc J. Van Gool, A. Oosterlinck · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1995
This paper describes a complete recognition system for planar objects, using semi-local projectively invariant information on the contours to be identified. The system combines two different canonical descriptions of these segments. Using semi-local information makes it possible to recognize objects which are only partially visible. This allows for realistic imaging situations. By doing recognition up to only a projective transformation, nothing has to be known about the camera position or focal length, which makes the system applicable in a wide range of applications. Classification of training data is done by BUCA, a new supervised classification algorithm which is both fast and reliable. BUCA incorporates the idea of the `minimal description length' principle by the use of splitting entropy. The final assembling of recognized segments tries to avoid polynomial running time by first combining the most certain matches, and subsequently eliminating all segments that are compatible with such a match.