Computer Perception of Curved Objects Using a Television Camera

Abstract

Various techniques are described for the computer perception of curved objects (cups, mugs, toruses, etc.). Research has been conducted in the areas of image-processing, object recognition, 3-d analysis, and scene analysis. A representation of images in terms of lines is argued to be superior to one in terms of regions. Line-finding is accomplished by an edge-follower which can track round curved boundaries. A new method of segmentation coupled to an improved procedure for fitting conic sections is used to obtain a line-drawing from the object boundaries. A topological description of the image is built up in which junctions are classified according to a comprehensive scheme applicable to pictures with both curved and straight lines.

'The performance of Barrow & Popplestone's program for recognising irregular objects is evaluated, and it is demonstrated that considerable improvements in speed. may be obtained by letting gross features of an object, such as its outline, direct the matching. Objects may also be identified with a new form of Waltz's techniques, based on labelling-constraints derived from topological, 3-d object models. It is shown how the hierarchical synthesis method for object recognition may be implemented so as to facilitate flexible interaction. Good tolerance of imperfections and rapidity of matching are achieved with this technique.

A qualitative measure of the shape of object surfaces is obtained from examination of the intensity contours created by the shading of reflected light. 3-d information is also acquired by matching image descriptions with "procedural" models of a set of prototype objects.

Scene analysis is performed by generalisations of the methods devised by Waltz. Description and partition of a scene is carried out using either models of specific objects or models of object classes. Procedures are explained for generating labels for the junctions caused by curved objects; the labels for a variety of types of surface interactions are tabulated in an appendix. A complete scene analysis system is described which integrates the research efforts in these areas. Its performance on simple scenes containing both curved and polyhedral objects is assessed. The thesis concludes with an appraisal of the results achieved, with particular regard to future lines of development.