Small UAS Landing Site Detection with ArUco Markers and Deep Learning Based Computer Vision

Small unmanned aerial systems (UAS) have shown promises in commercial package delivery services. Numerous studies have shown ArUco markers to be an effective and robust method to enhance the identification of landing sites during UAS descent. Practical deployment of these markers at warehouse landing sites (such as a Walmart parking lot), where the UAS can land and load packages for delivery, is feasible. However, it seems an unlikely assumption that a customer or end-user would be able to set up an ArUco marker landing configuration at their home delivery site, such as a front yard, backyard, or on top of the apartment building with the correct ordering and proper dimensions. It is much more user-friendly to require only a generic landing pad that the UAS can recognize using deep learning techniques; Thus, this paper investigates the feasibility of using deep learning based computer vision models to recognize the landing pad without relying on ArUco markers. To that end, we describe a custom data collection methodology using flight videos featuring a landing pad, along with the dataset annotation procedure in Roboflow. This annotated dataset is used to train multiple YOLOv8 object detection models, each initialized with different weights, to accurately detect the landing site. These models are integrated into a proposed Obstacle and Landing Site Awareness System (OLSAS) designed to support safe and autonomous UAS package delivery landings. Finally, we compare the detection performance and reliability of the deep learning-based approach against an ArUco marker detection algorithm to evaluate their effectiveness in identifying landing sites.