Research On The Principles And Algorithms For Recognizing Fully Or Partially Occluded Object

As an interesting research in computer vision, the object recognition has been found wide applications in various technical sections, such as robot automatic picking, automatic navigating, and detecting object automatically, assembling automatically, analyzing medical and remote sensing image, and so on. Three-dimensional information of an object will be lost while it is projected into a two-dimensional image, so the purpose of recognition is to identify three-dimensional object based on the hidden information in two-dimensional images.This dissertation is mainly to investigate the principles and algorithms for recognizing two-dimensional planar objects, three-dimensional polyhedrons and curved surface objects from a fully or partially occluded line drawing or image.The definitions of homograph are given under axonometric and perspective projections in terms of topological and geometrical characteristics, which are employed to describe polygons. Based on the principle of homograph, a new method is proposed for recognizing that two polygons are homographs. On the basis of recognizing homographs, new algorithms are also proposed for recognizing polyhedrons from an axonometric or perspective line drawing and image. The experiment results show that these algorithms can not only recognize polyhedrons of different shape but also distinguish between polyhedrons that are of the same shape but their sizes and proportions are different. Comparing with other recognizing algorithms, the algorithms presented in this dissertation have improved obviously.A new method is proposed for recognizing partially occluded polygon objects under affine transformation. First, a local invariant is given under affine transformation, and a normalized similarity function is established. And then, a function of loss features is constructed to judge whether each local feather is lost or not, and similarity measure is calculated only using the local features that are not lost, so polygon objects can be recognized from a partially occluded line drawing or image. Noise and occlusion are analyzed in constructing the normalized similarity function and the judge function of lost feathers. The experiment results show that the proposed recognizing algorithms are insensitive to noises and occlusion.A new method is put forward to match two-dimensional curves under affine transformation. The definition of NRLCTI (Normalized Run Length Code of Conner and Tangent and Inflexion Points) of a two-dimensional curve is given. In terms of NRLCTI, we can match feature points both on object and models preliminarily. A method for estimating optimal affine transformation is given based on Frobenius norm. A new algorithm is designed to match sub-curves, which can cope with the problem that the curve represented by the feature points isn't unique. A novel approach is set up to recognize curves from a line drawing or image. By partitioning the curve into many sub-curve based on landmarks, then matching and recognizing them, the low accuracy for curve approximated by polygon can be overcome. The experiment results are also given.