Sources of Uncertainty in Feature-Based Image Registration Algorithms

One significant technological barrier to enabling multi-sensor integrated ISR is obtaining an accurate understanding of the uncertainty present from each sensor. Once the uncertainty is known, data fusion, cross-cueing, and other exploitation algorithms can be performed. However, these algorithms depend on the availability of accurate uncertainty information from each sensor. In many traditional systems (e.g., a GPS/IMU-based navigation system), the uncertainty values for any estimate can be derived by carefully observing or characterizing the uncertainty of its inputs and then propagating that uncertainty through the estimation system. In this paper, we demonstrate that image registration uncertainty, on the other hand, cannot be characterized in this fashion. Much of the uncertainty in the output of a registration algorithm is due to not only the sensors used to collect the data, but also data collected and the algorithms used. In this paper, we present results of an analysis of feature-based image registration uncertainty. We make use of Monte Carlo analysis to investigate the errors present in an image registration algorithm. We demonstrate that the classical methods of propagating uncertainty from the inputs to the outputs yields significant underestimates of the true uncertainty on the output. We then describe at least two possible sources of additional error present in feature-based methods and demonstrate the importance of these sources of error.