Detecting infrared UAVs on edge devices through lightweight instance segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Infrared UAV detection for surveillance requires balancing accuracy, real-time processing, and edge deployment capabilities.
• Existing deep learning models often compromise one of these critical requirements.

Purpose of the Study:

• To present YOLO11-AU-IR, a novel lightweight instance segmentation framework for infrared UAV detection.
• To address the conflicting demands of accuracy, speed, and resource efficiency in surveillance applications.

Main Methods:

• Introduced Efficient Adaptive Downsampling (EADown) to preserve small-target features during multi-scale fusion.
• Developed HeteroScale Attention Network (HSAN) for enhanced cross-scale feature representation using grouped convolutions and attention mechanisms.
• Implemented Adaptive Threshold Focal Loss (ATFL) with epoch-adaptive tuning to handle foreground-background imbalance in infrared imagery.

Main Results:

• Achieved 97.7% mAP@0.50 and 75.2% mAP@0.50:0.95 on the AUVD-Seg300 dataset, outperforming baselines.
• Reduced parameters by 24.5% and GFLOPs by 11.8% while maintaining 59.8 FPS inference.
• Demonstrated practical edge compatibility on NVIDIA Jetson TX2 with 95% mAP@0.50 under INT8 CPU-only deployment.

Conclusions:

• YOLO11-AU-IR effectively balances accuracy, real-time performance, and edge deployment for infrared UAV detection.
• The architectural innovations (EADown, HSAN, ATFL) significantly contribute to the model's efficiency and robustness.
• Publicly available code and dataset facilitate further research and application in surveillance.