Enhancing two-stage object detection models via data-driven anchor box optimization in UAV-based maritime SAR.

Area of Science:

• Computer Vision
• Robotics
• Maritime Safety

Background:

• Unmanned Aerial Vehicles (UAVs) are vital for high-altitude maritime Search and Rescue (SAR).
• Accurate object detection in UAV imagery is crucial for identifying boats, personnel, and objects.
• General object detection models struggle with the unique challenges of maritime SAR imagery.

Purpose of the Study:

• To analyze the unique attributes of UAV-based maritime SAR image data.
• To identify the need for optimizing object detection for difficult-to-detect targets in this domain.
• To propose and validate an anchor box optimization strategy for two-stage object detection models.

Main Methods:

• Leveraged the SeaDronesSee benchmark dataset for UAV-based maritime SAR.
• Employed clustering analysis to develop an anchor box optimization strategy.
• Conducted experiments to validate the proposed method and analyze its effectiveness.

Main Results:

• Achieved a 45.8% increase in average precision over default torchvision anchor boxes.
• Demonstrated a 10% increase in average precision over SeaDronesSee official sample code configurations.
• Significantly improved detection of swimmers, floaters, and life jackets on boats.

Conclusions:

• The proposed anchor box optimization strategy enhances object detection performance in UAV-based maritime SAR.
• This method offers valuable insights into data characteristics and model optimization for the research community.
• Provides a meaningful reference for future advancements in maritime SAR technology.