Advancing Autonomous Aerial Refueling with Deep-Neural-Network-Based Image Segmentation

Integrating uncrewed aerial systems (UASs) into United States Navy carrier air wings will most likely require the implementation of autonomous aerial refueling (AAR). AAR will boost operational flexibility, range, and endurance. However, the current challenge lies in achieving the necessary safety of flight certification for AAR. This study introduces a deep neural network model trained with a limited dataset from actual flight test footage to accurately identify and track segments of the refueling drogue. Unlike traditional object detection methods, this model employs instance segmentation to dynamically outline the drogue’s shape. This advancement lays the groundwork for future models capable of predicting the drogue’s shape and position, even under challenging environmental conditions and potential false detections. The model’s efficacy is evaluated using video footage from KC-130 refueling operations, encompassing diverse background scenarios and noise levels. The model demonstrates robust performance, achieving a high mean average precision of 0.971 at an intersection over union threshold of 0.5. By addressing real-world challenges of image quality and varying light conditions, this research aims to significantly advance the reliability and sophistication of AAR technologies.