Using Buildings Proximity Information in UAV Global Localization Over Satellite Images

Unmanned Aerial Vehicles (UAVs) depend on localization systems to safely and efficiently perform their tasks, usually relying on GPS. However, this setup is not problem-free, and having only one source of measurement may be catastrophic to the UAVs' operation. To avoid this kind of issue, this paper proposes a visual UAV localization system as an extra source of pose estimation in case of GPS malfunctioning. The novelty of our system is its measurement model used in the Monte Carlo Localization, which relies on a new descriptor called NBD-BRIEF, based on the Nearest Building Distance (NBD) information obtained at the vehicle position. The building footprint is extracted from the UAV images using a convolutional network. Such semantic information is more stable and less sensitive to light and color changes, making the method generally more robust than approaches purely based on color matching. The experiments have shown that our NBD-BRIEF descriptor performs better than competing approaches to the same problem. And the results indicate that our visual UAV localization system based on NBD-BRIEF can properly estimate the UAV's pose in three different flights, while other localization methods we compared failed.