Heuristic data-driven anchor generation for UAV-based maritime rescue image object detection.

Area of Science:

• Computer Vision and Image Analysis
• Robotics and Autonomous Systems
• Maritime Safety and Rescue Technologies

Background:

• Maritime rescue operations using Unmanned Aerial Vehicles (UAVs) face challenges in accurately detecting targets in complex visual data.
• Current object detection models often struggle with the specific scenarios and tasks encountered in drone-based maritime surveillance.

Purpose of the Study:

• To enhance the performance of object detection models for UAV-based maritime rescue by optimizing anchor box generation.
• To investigate heuristic methods for extracting relevant data features from UAV maritime rescue imagery.

Main Methods:

• Utilized heuristic methods to extract data features from UAV maritime rescue images.
• Optimized anchor box generation within the MMDetection object detection framework.
• Conducted experiments on the large-scale SeaDronesSee maritime rescue dataset.

Main Results:

• Optimized anchor boxes improved model performance by 48.9% to 62.8% compared to default configurations.
• The most proficient model exceeded the official SeaDronesSee baseline performance by over 49.3%.
• Analysis identified variations in detection difficulty for different objects and explained underlying reasons.

Conclusions:

• The proposed heuristic method for anchor box optimization significantly improves UAV maritime rescue object detection.
• The findings offer a promising approach for refining data analysis and enhancing maritime rescue capabilities.
• This research contributes to the advancement of autonomous systems in critical search and rescue applications.