Season-Invariant GNSS-Denied Visual Localization for UAVs

Localization without Global Navigation Satellite Systems (GNSS) is a critical functionality in autonomous operations of unmanned aerial vehicles (UAVs). Vision-based localization on known map can be an effective solution, but it burdened by two main problems: places have different appearance depending weather and season, the perspective discrepancy between UAV camera image make matching hard. In this work, we propose solution relying images to georeferenced orthophotos with trained convolutional neural network model that invariant significant seasonal difference (winter-summer) map. We compare convergence speed accuracy our six reference methods. The results show major improvements respect methods, especially under high variation. finally demonstrate ability method successfully localize real UAV, showing proposed robust changes.