Multi-Scale Oriented Detection with Shared Convolution for UAV-Enabled Maritime Safety Surveillance.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Maritime Surveillance

Background:

• Unmanned aerial vehicles (UAVs) with high-definition cameras are vital for maritime safety management due to their wide coverage and adaptable perspectives.
• Challenges in UAV-based maritime surveillance include inaccurate detection across varied angles and scales, and computational strain from large models.

Purpose of the Study:

• To develop a lightweight, multi-scale oriented detection model specifically designed for UAVs to address challenges in maritime safety surveillance.
• To improve the accuracy and efficiency of vessel detection in complex maritime environments using UAV imagery.

Main Methods:

• Proposed a novel LDFusion module for cross-stage feature fusion, enabling adaptable scale extraction to accommodate variable flight altitudes.
• Designed a lightweight detection head with shared convolution modules (SConvs) to reduce model parameters for oriented ship detection.
• Created three oriented datasets from a maritime UAV perspective, including new and re-annotated inland and marine datasets.

Main Results:

• The proposed lightweight model achieved a modest 3.27% improvement in detection accuracy.
• Significantly reduced the number of parameters by 24.40% compared to state-of-the-art approaches.
• Demonstrated effective performance across three diverse maritime UAV datasets.

Conclusions:

• The developed lightweight multi-scale oriented detection model offers a viable solution for efficient and accurate maritime surveillance using UAVs.
• The model effectively balances detection performance with reduced computational complexity, making it suitable for resource-constrained UAV platforms.
• The creation of new datasets facilitates further research and development in UAV-based maritime safety applications.