An Analysis of Multi-Object Detection on 2024 Aerial Refueling Flight Test Data

Autonomous aerial refueling (AAR) enables extended mission endurance for both manned and unmanned aircraft without relying on ground-based support. Traditional AAR methods use active sensors such as radar and light detection and ranging (LiDAR), which are susceptible to jamming and interference. This paper investigates a passive, vision-based approach using You Only Look Once (YOLO) convolutional neural networks (CNNs) to detect and track refueling components using only monocular imagery. Two flight test campaigns were conducted to evaluate system performance: one using a scalemodel aircraft and the other using full-scale aircraft. The results demonstrate that larger objects like tankers can be accurately tracked at greater distances, while smaller objects like drogues require closer proximity for reliable pose estimation. A method leveraging a separate network for tanker tracking provides azimuth and elevation cues, guiding the receiver toward the drogue until it is close enough for precise docking. Additionally, extensive flight imagery was collected for future validation using recorded Global Positioning System (GPS) data. These findings highlight the feasibility of vision-based AAR and inform future work toward a fully autonomous refueling capability.