Real-Time Object Detection for Autonomous Solar Farm Inspection via UAVs

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Solar farm inspection requires agile and precise object detection.
• Conventional methods like bounding boxes or segmentation lack granularity for detailed inspection.
• Unmanned aerial vehicles (UAVs) are increasingly used for industrial asset inspection.

Purpose of the Study:

• To introduce an innovative keypoint-based object detection framework for real-time solar farm inspections.
• To improve the granularity of object detection by focusing on solar panel vertices.
• To optimize the framework for embedded platforms for efficient robotic operations.

Main Methods:

• Developed a keypoint-based object detection framework inspired by CenterNet.
• Optimized the architecture for embedded platforms (e.g., NVIDIA AGX Jetson Orin).
• Integrated active learning strategies to reduce annotation efforts.

Main Results:

• Achieved near 60 FPS at 1024 ×1376 resolution, exceeding camera operational frequency.
• Demonstrated a real-time capability essential for time-critical industrial inspections.
• Model design emphasizes reduced computational demand for practical deployment.

Conclusions:

• Keypoint-based object detection offers a practical and effective approach for UAV-based solar farm inspections.
• The proposed framework provides richer granularity compared to traditional methods.
• The system is optimized for real-time performance and reduced computational load.