dc.description.abstract | In recent years the use of unmanned air systems (UAS) has seen extreme growth. These small, often inexpensive platforms have been used to aid in tasks such as search and rescue, medicinal deliveries, disaster relief and more. In many use cases UAS work alongside unmanned ground vehicles (UGVs) to complete autonomous tasks. For end-to-end autonomous cooperation, the UAS needs to be able to autonomously take off and land on the UGV. Current autonomous landing solutions often use fiducial markers to aid in localizing the UGV relative to the UAS, an external ground computer to aid in computation, or gimbaled cameras on-board the UAS. This thesis seeks to demonstrate a vision-based autonomous landing system that does not rely on the use of fiducial markers, completes all computations on-board the UAS, and uses a fixed, non-gimbaled camera. Algorithms are tailored towards low size, weight, and power constraints as all compute and sensing components weigh less than 100 grams. The foundation of this thesis extends upon current efforts by localizing the UGV relative to the UAS using neural network object detection and camera intrinsic properties instead of common place fiducial markers. An object detection neural network is used to detect the UGV within an image captured by the camera on-board the UAS. Then a localization algorithm utilizes the UGV’s pixel position within the image to estimate the UGV’s position relative to the UAS. This estimated position of the UGV will be passed into a command generator that sends setpoints to the on-board PX4 flight control unit (FCU). This autonomous landing system was developed and validated within a high-fidelity simulation environment before conducting outdoor experiments. | |