On a mathematical framework for object recognition from multi-perspective remotely sensed imagery

We develop a new perspective invariant feature space representation of remotely sensed objects, regarding the features themselves as primitive observables of the 3D objects and to estimate them from multiple sensor measurements. This is formulated as an inverse problem in the feature coefficients. Once the coefficients are estimated they may be used to derive higher level features used by machine learning algorithms for classification. The focus of this paper is on the mathematical formulation of the feature estimation problem from one or more perspective images. We also give a discussion of how this fits into a larger object classification system.