Discriminating Similar Naval Vessels Using YOLOv8 Deep Neural Network

Abstract

The next generation of United States Navy uncrewed aerial systems (UASs) is expected to operate in global positioning system and radio frequency-denied maritime environments. In these challenging conditions, these UASs must accurately identify specific surface vessels among multiple similar vessels using passive onboard sensors. This study explores the potential of a deep neural network (DNN) to differentiate between three similar surface vessel classes using actual footage of the vessels underway within their operational environments. The DNN’s effectiveness is evaluated using data collected under diverse environmental conditions, including different times of the day and various sky conditions, which imply varying levels of light and visibility. The trained DNN model demonstrated outstanding performance on real-world maritime datasets, achieving a mean Average Precision of 94.2% at an intersection over union of 0.5, effectively distinguishing vessels with minimal false positives. Our findings demonstrate that, with proper training, a DNN model can accurately differentiate between vessels despite their similarity and under challenging conditions.

Keywords

• Deep neural networks (DNN)
• military operations
• GPS/RF-denied environments
• maritime operations
• surface vessel identification
• uncrewed aerial systems (UASs)