A Study on Real-Time Safety Equipment Detection Model based on Skeleton-Fused Technique in Embedded Environments

Key Points

• The aim is to develop a real-time safety equipment detection model that integrates skeleton-fused techniques for better accuracy.
• Utilized YOLOv11n as a lightweight backbone model.
• Combined skeleton extraction information for enhanced detection.
• Applied logic to reduce false positives by verifying object and joint overlap.
• Evaluated model performance using mAP50 metrics.
• Achieved a mean Average Precision (mAP50) of 57.21%, improving by 5.3% over the previous single model.
• Significantly enhanced the detection accuracy of small objects, such as goggles, by 37.4%.
• Maintained a stable inference speed of 20.2 FPS, enabling real-time safety management.

Abstract