Three dimensional recognition from depth images by detailed surface encoding (moment invariants)
Richard Poulo · 1985
This thesis describes a new approach to object recognition in three dimensions. The problem of three dimensional recognition is often reduced to the two subproblems of recognizing object components and recognizing the spatial relations between the components. Most previous work has concentrated on recognizing the spatial relations between components while approximating the components with easily parameterized surfaces. This emphasis has been a matter of necessity rather than of choice because it was difficult or impossible to parameterize the actual variations of contoured surfaces. The approach in this thesis is to concentrate on the 3D details of object structure, where it is assumed that objects to be recognized have already been isolated from the remainder of the scene they appear in. This approach is based on using numeric features that remain invariant under a general rotation in 3-space for a fixed visible part of an object. The features are derived from two dimensional moment invariants using the density function in the moments to encode the deviations of a contoured three dimensional surface from a plane. Hidden or partially hidden object surfaces are handled by segmenting objects into their distinct faces and applying the recognition features to each face. A new technique for segmenting an object into its faces and the recognition procedures for both the individual faces and the entire object are discussed. While the results of an actual implementation demonstrate that these features alone suffice to implement a fast, reliable object recognition procedure it is expected that the present approach of concentrating on the detailed contouring of 3D objects will eventually complement techniques relying on component spatial relations to yield more powerful recognition systems than now exist.