An improved algorithm for infrared road object recognition in UAV perspective.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Unmanned aerial vehicles (UAVs) are crucial for ground-object recognition, but infrared (IR) imagery presents challenges like low contrast and background clutter.
• Existing methods struggle with detecting small targets in complex thermal scenes.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for infrared object detection from UAVs.
• To address limitations in contrast, clutter, and target size in thermal imagery.

Main Methods:

• Introduced YOLO-IR, a lightweight detector based on YOLOv7, incorporating a global-efficient backbone for enhanced thermal-texture modeling.
• Integrated parameter-free SimAM attention for salient IR structure highlighting and an efficient BiFPN for multi-scale fusion.
• Utilized normalized Wasserstein distance for scale-insensitive localization across assignment, regression, and non-maximum suppression.

Main Results:

• YOLO-IR achieved 94.5% precision, 92.9% recall, and 95.7% mAP@0.5 on a UAV thermal dataset.
• Demonstrated significant improvements over the YOLOv7 baseline (+4.3% P, +1.8% R, +4.2% mAP) while maintaining real-time performance.
• Qualitative results showed reduced misses and false alarms in dense, low-contrast scenes.

Conclusions:

• YOLO-IR provides accurate and efficient infrared road-object recognition from UAV viewpoints.
• Each component of YOLO-IR contributes to consistent performance gains.
• The model effectively handles challenges posed by adverse illumination and complex thermal scenes.