Blob-Geometry Monocular Depth Estimation for Drogue Tracking in Autonomous Aerial Refueling

Reliable monocular depth perception is essential for autonomous probe-and-drogue aerial refueling, particularly in GPS-denied and/or communication-constrained environments. We introduce a lightweight depth-estimation framework that detects the drogue using a YOLO-based segmentation front end. The model fits an ellipse to its visible rim and computes depth analytically from blob-level geometric features using the pinhole camera model. These features include the ellipse’s area, major and minor axes, eccentricity, and in-plane rotation, which serve as robust geometric cues under perspective projection. Depth is inferred directly from the recovered ellipse major axis and the known drogue diameter, eliminating the need for learned depth regressors and improving resilience to occlusion, motion blur, and off-axis viewpoints. Ground-truth measurements from a 16-camera OptiTrack system enable precise benchmarking across the full operational envelope. The proposed segmentation–geometry method achieves median depth errors below approximately 2 in and maintains RMSE under about 4 in for nearly the entire 7.5–22.2 ft envelope. Additionally, the model exhibits a mild, consistent underestimation bias with tightly bounded error distributions. These results demonstrate that the approach delivers accurate, stable, and real-time monocular ranging suitable for embedded perception modules aboard uncrewed aerial systems.