Information Fusion Using N-Dimensional Hashing
Abstract
During Phase 1, I-MATH Associates, Inc. and the NYU Courant Institute of Mathematical Sciences have developed algorithms and real-time software for fusion of 3D imagery and information. The fundamental technique is geometric hashing. Hashing is an efficient method for storing a very large set of models, representing various target types and poses, and then quickly determining which model best represents an unknown item, whose corresponding features are sifted through the hash table. In its current form, hashing represents an object's (or scene's) feature values in a 2D table whose abscissa and ordinate correspond to the feature variables. Typically, such features are (x,y) geometric coordinates of key interest points about the object (scene). However, the features can be any basis function, including affine transforms of a rigid body, radius of curvature and tangent magnitude of curved objects, etc. Hence, hashing allows disparate types of information to be placed in a common table. The overall objective of this STTR is not just multidimensional pattern recognition, but rather maximum extraction of information from multiple sources, which may be dissimilar and perhaps not even imaging. Hashing directly supports such fusion, since multiple types of features can be the basis for an nD hash table. The thrust of this STTR development has been to devise an nth order hashing schema, beginning with a 3D implementation for Phase 1. However, our approach is not limited to extending the hash table from a 2D to 3D (or higher dimension) domain. We have also investigated alternative techniques during Phase 1, including: (1) Hashing on 2D plane orthonormal projections, and then combining the results using postclassifier fusion techniques
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 08, 1997
- Accession Number
- ADA332199
Entities
People
- Harley Myler
- Ronald Patton