Cramer-Rao bound on projective image registration from feature points

We derive Cramer-Rao lower bounds (CRLBs) on the estimation of projective image registration parameters from control points. We also bound the variance of the estimated position of a transformed pixel, and we extend this bound to an estimate of the second-order statistics of the pixel intensities. The variance of the pixel intensities varies spatially, and large variances occur at locations for which we have low confidence in the accuracy of the registered pixel intensities. Such information may be used in exploitation tasks to improve decision accuracy or for decision confidence reporting.