Real-time aerodrome detection using deep learning methods

Abstract

Locating aerodromes from a large distance or in low visibility can be difficult for pilots when they experience spatial disorientation. Therefore, this research creates an aid for pilots to increase their spatial awareness by introducing a novel technology for visual detection of aerodromes from an aircraft in real-time. Aerodromes differ significantly in size, layout, number of runways, and other factors. Moreover, their visual appearance changes with altitude, range, and weather conditions. To address these numerous combinations, a comprehensive solution is proposed which uses the state-of-the-art YOLOv5 object detection algorithm. The YOLO model was trained with real and synthetic data from visible imaging sensors, as well as data augmentation techniques, to increase performance for low visibility conditions. An image type analysis showed that satellite images are not useful for detecting aerodromes from an aircraft due to perspective differences. Single- and multi-class detection were compared and advantages were identified for detecting runways in addition to aerodromes. The final performance to detect aerodromes is over 0.9 mAP@0.5 in good visibility and does not drop below 0.6 in low visibility, with an execution speed of more than 60 fps. This final model does not depend on other on-board systems and is capable of detecting aerodromes at different altitudes and distances from an aerodrome, in low visibility, regardless of camera position/orientation and regardless of other objects in the frame.