Previous Abstract Return to Session E9 Next Abstract

Session E9: PNT for Autonomous Systems

Physics-Informed Computer Vision for Long-Range Aerial Refueling
Nathaniel Vinoya, Jiwoon Yim, Scott Nykl, AFIT
Location: Ballroom B
Alternate Number 1

Autonomous visual relative navigation to refueling aircraft in GPS-denied environments is a critical facilitator for Autonomous Aerial Refueling (AAR). However, when implementing a fixed sensor and lens on the receiver aircraft, the refueler aircraft’s projected resolution shrinks inversely with distance, forcing its projected pixel density to fall below the minimum pixel quality of standard Convolutional Neural Networks (CNNs). This research addresses the domain beyond 3km, where You-Only-Look-Once (YOLO) models are unable to perceive refueler aircraft projections due to their small size. Ultimately, this constraint exposes the Downsampling Paradox, causing spatial detection failures even when paired with high-resolution sensors.
A recent study showed that a YOLOv10 network reliably detects aircraft only up to 2.2km. Since details are lost past that distance, the target renders a featureless blob that is indistinguishable from clouds, birds, or other noise in the sky, necessitating a shift from appearance-based detection to motion analysis. Standard tracking methods, such as linear Kalman filters, fail in this task because they assume predictable, straight-line motion. However, receiver aircraft often banks or maneuver during refueling operations, creating non-linear motion patterns in the background optical flow that cause standard filters to fail. Thus, a Heuristic Blob Detector first flags all potential objects in the frame based on basic visual features. To isolate the true target, a Physics-Informed Multi-Hypothesis tracker analyzes candidates by their movements; by filtering specific flight behaviors during refueling operations, such as velocity consistency and trajectory smoothness, the tracker distinguishes the refueler aircraft from atmospheric noise.
Validation in the AftrBurner engine revealed a decisive performance crossover point at approximately 2.2km, marking the range where an F1-score of 0.94 is achieved by the proposed system. This outperformed the YOLO baseline F1-score of 0.64 in the feature-scarce regime. Current research investigates transferring physics-based scoring to real-world flight footage. These tests aim to establish that classical heuristics, when augmented by physics-aware state estimation, provide a reliable solution for feature-scarce tasks.
Key Contributions:
1. Analysis of the Downsampling Paradox, explaining why CNNs fail in long-range AAR when the projected target falls below the minimum pixel size.
2. Development of a Physics-Informed tracker that replaces learned appearance-based detection with motion analysis to maintain track continuity.
3. Integration of background optical flow filtering to distinguish between static cloud clutter and moving targets during the receiver aircraft’s non-linear banking maneuvers.

Previous Abstract Return to Session E9 Next Abstract