Registration Of Forward Looking Imagery

Technical Abstract We propose to develop a fast technique for perfing a robust, single-frame registration of images captured with forward looking cameras or other sensors. Unlike conventional techniques, which require that a large number of reference templates be compared with a sensor image, the proposed technique will directly determine the affine transformation (i.e., translation, scaling, and rotation) necessary to register the sensor and reference (template) images. The proposed technique will directly determine the six parameters that represent the required affine transformation. The technique will also determine the two additional parameters required to specify a complete perspective transformation. We will also evaluate techniques for computing more general elastic transformations that may be required to fit a template to a sensor image that has been distorted nonlinearly. The novel solutions to robust image registration proposed here will be compared to more conventional image registration techniques. Finally, a prototype software system will be implemented on a general-purpose workstation in order to demonstrate the feasibility of the system proposed for development in Phase II.