2-D object recognition using invariant contour descriptor and projective refinement

摘要

This paper presents an efficient model-based recognition method to recognize 2-D objects and to obtain correspondences between models and scene boundaries with a subpixel positioning error.As a shape signature for a contour, we propose a descriptor consisting of five-point invariants, which are used to index a hash table. Also, we propose a projective refinement as a verification method to compute exact correspondences between models and scene contour points. This method repeatedly computes projective transformation using a weighted pseudo inverse.We present an error model for five-point invariants, which are used to define a similarity between two descriptors, to determine a searching bound in indexing, and to obtain the weights in the projective refinement.In experiments using seriously distorted images of forty models, this method led to the recognition of planar curved objects. A transformation using the correspondence between the model and scene points on contours was also obtained.