Visual Localization for UAVs in Outdoor GPS-denied Environments

Vision-based navigation techniques are commonly used for autonomous flight in outdoor GPS-denied environments. We adapt Visual Teach and Repeat (VT), a vision-based autonomous route-following system, to use on multirotor UAVs. Since multirotors are underactuated, there are no guarantees that the camera viewpoint will be the same at matching positions along the teach and repeat paths, which causes visual localization to be challenging. The first part of this thesis demonstrates that by using a 3-axis gimballed camera with an appropriate active pointing strategy, we can improve the visual localization performance and robustness within the VT framework. The second part of this thesis presents a method to estimate the global pose of a UAV by using an information-theoretic approach to register real images with rendered georeferenced images from 3D Google Earth. We show that this method is capable of accurately estimating the 6DoF pose with a position accuracy on par with GPS.M.A.S