3-D object recognition and orientation from single noisy 2-D images

We have developed a method of recognising 3-D objects and determining their orientation from single 2-D images. These images may be noisy and contain partially occluded shapes. The method combines both analytical and structural type techniques to build a description of the viewed object. This description is then compared to a database of 2-D model descriptions for various object orientations using 2-D pose-clustering techniques. Features derived from the local geometry of the object boundary are used in the matching process. This enables the object to be recognised and its orientation determined to within an average uncertainty of less than six degrees of are, in both the major and minor axes of the object.