Deep Learning-Based Relative Bearing Estimation Between Naval Surface Vessels and Uas in Challenging Maritime Environments

This paper introduces a deep neural network (DNN) framework designed to accurately determine the relative bearing between a naval surface vessel and an uncrewed aerial system (UAS) using video data collected from drone operations in a maritime environment. Utilizing a dataset of 2,773 augmented images categorized into twelve 30-degree angular classes, the DNN was trained with the YOLOv11s architecture to identify and localize the relative bearing of a Yard Patrol vessel effectively. The model demonstrated exceptional performance, achieving an overall precision of 89 %, recall of 91.3 %, and a mean average precision (mAP) of 93.6 % at a 50 % intersection over union (IoU) threshold. Furthermore, the mAP averaged 83.1% across IoU thresholds from 50% to 95%, highlighting the model's robustness in diverse conditions. These metrics indicate that the DNN can reliably estimate the UAS's bearing relative to the naval vessel, thereby facilitating autonomous recovery operations in communication-denied maritime environments. This advancement supports the United States Navy's initiative to integrate rotary-wing UASs onto existing naval platforms, ensuring safe and efficient flight operations from stern landing areas.