Uncrewed Aerial System Vision-Based Relative Bearing Estimation for Maritime Recovery
TL;DR
A deep neural network framework designed specifically to address this gap is presented, detailing its validation from a laboratory proof-of-concept to at-sea trials that validate that the DNN can reliably estimate a UAS’s bearing relative to a naval vessel.
Abstract
Accurate relative positioning of uncrewed aerial systems (UAS) is paramount for safe autonomous operations, particularly in maritime environments where communication may be denied. While much of the existing research on vision-based ship landing focuses on solving for depth or full six-degree-of-freedom pose, often with the aid of specific markers, robust 360 deg relative bearing estimation remains a critical and comparatively underexplored component for precise navigation and landing maneuvers. This paper presents a deep neural network (DNN) framework designed specifically to address this gap, detailing its validation from a laboratory proof-of-concept to at-sea trials. The laboratory experiment demonstrated that a DNN could correctly identify the relative position of the UAS from the vessel using a YOLOv8 architecture. For the at-sea analysis, a YOLOv11 architecture was trained on a data set consisting of actual and augmented images. The resulting model demonstrated exceptional performance, achieving a mean average precision of 96.8% at an intersection over union (IoU) threshold of 50% and 86.0% across IoU thresholds from 50 to 95%, with overall 95% precision and 94% recall. These robust metrics validate that the DNN can reliably estimate a UAS’s bearing relative to a naval vessel, providing a critical capability for enabling safe and efficient autonomous recovery operations.