Deep Learning-Based Super-Resolution Reconstruction and Marker Detection for Drone Landing

Abstract

There have been a significant number of recent studies on autonomous landing in unmanned aerial vehicles (UAVs). Early studies employed a global positioning system (GPS) receivers for this purpose. However, because GPS signals cannot be used in certain urban environments, prior studies used vision-based marker detection. To accurately detect a marker, a high-resolution camera on a drone must obtain a high-quality image. This can not only be expensive but also increases the weight of the drone. In general, drones are only equipped with a frontal-viewing and fixed angle camera, and an additional downward-viewing camera becomes necessary for drone landing. Therefore, expensive and weighted high-resolution cameras are not feasible for use on drones. Nevertheless, most previous studies on vision-based drone landing use high-resolution images. To address such limitations, we propose a new method of drone landing using deep learning-based super-resolution reconstruction and marker detection on an image captured by a cost-effective and low-resolution visible light camera. The experimental results on two datasets demonstrate that our method exhibits higher performance than the existing methods in terms of super-resolution reconstruction and marker detection.

Keywords

• Unmanned aerial vehicle
• autonomous landing
• deep learning-based super-resolution reconstruction
• deep learning-based marker detection
• visible light camera on drone