Improving Object Detection in High-Altitude Infrared Thermal Images Using Magnitude-Based Pruning and Non-Maximum...

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• High-altitude platforms, particularly unmanned aerial vehicles (UAVs), offer advantages for remote sensing.
• Infrared thermal object detection from elevated positions captures unique thermal signatures.
• Processing high-resolution thermal imagery presents challenges for traditional detection methods.

Purpose of the Study:

• To explore the application of YOLOv8 architecture for high-altitude infrared thermal object detection using UAVs.
• To enhance object detection speed and accuracy in thermal imagery.
• To address the complexities of processing high-resolution thermal data.

Main Methods:

• Utilized YOLOv8's advanced architecture combined with dynamic magnitude-based pruning and non-maximum suppression.
• Converted COCO and PASCAL VOC dataset annotations to YOLO format for efficient training and inference.
• Applied techniques to high-altitude infrared thermal imagery captured by UAVs.

Main Results:

• The proposed YOLOv8 architecture demonstrated superior speed and accuracy in thermal object detection.
• Precision-recall metrics confirmed robust performance in identifying objects based on thermal signatures.
• Some misclassification, especially for human subjects, was observed, indicating areas for future improvement.

Conclusions:

• YOLOv8's advanced architecture shows significant potential for improving UAV-based thermal imaging applications.
• The study paves the way for more effective real-time object detection solutions in extensive areas.
• Further refinement is needed to address specific misclassification challenges for enhanced overall performance.