The main objective of the research presented here is the minimization of the data set required for the reliable recognition of arbitrarily positioned three-dimensional objects. The analytical framework for studying 3-D object acquisition and recognition from 2-D images is based on invariant indexing. The objective is to identify a 3-D object from 2-D images, taken from arbitrary spatial points of view. The geometric relation between 3D-object and the resulting 2D image is modeled as an affine transformation. Objects are modeled as sets of characteristic spatial points, such as corner and edge points, which are stored in a hash table. In a new view, known objects are quickly identified, and their orientation is estimated as long as the viewing angle deviates not too much from the angle at which the model was generated. Quantitative analysis about the range of permissible viewing angle variations and confidence intervals are presented.
展开▼