Precise Estimation of Geo-location Uncertainty

ABSTRACT: Algorithms that perform accurate geo-registration of imagery captured from an aerial platform are becoming prevalent in the surveillance community. A universal shortcoming of these algorithms is their inability to provide an accurate estimate of the uncertainty of the solution. In this proposal we describe a solution for geo-registration based around the popular simultaneous localization and mapping (SLAM) approach. This solution includes a graphical model that captures the complex relationships between the input observations (2D keypoints extracted from the imagery and tracked frame-to-frame) and the hidden variables that are being estimated (the poses of the platform at each frame). The model also provides an explicit representation for the uncertainty in the poses of the aircraft and directly characterizes the noisy input observations as inliers or outliers. This model is combined with an efficient inference procedure that is able to overcome some inherent challenges to the problem, namely that the poses of the aircraft are not Euclidean (and therefore nontrivial to model) and the combinatorial nature of labeling each input as an inlier/outlier. We provide an outline for experiments on real and synthetic data and metrics to validate the accuracy of the uncertainty estimates from the algorithm. BENEFIT: The military currently employs geo-registration algorithms in a number of application areas (e.g. image-based navigation, targeting, situational awareness). Accurate uncertainty characterization in solutions from geo-registration algorithms will provide a measure of confidence in the result, and enables principled fusion of the resulting solution with measurements from other sensors/algorithms. This has direct impact on image-based navigation, as mentioned above, where the geo-registration solution could be used to aid the system in GPS-denied environments. Because the approach is an extension to the SLAM algorithm, it has direct commercial application to virtually any autonomous robotic deployment scenario, like emergency search and rescue and remote exploration.