A method for detecting safety helmets underground based on the YOLOv11-SRA model

Area of Science:

• Computer Vision
• Industrial Safety
• Deep Learning

Background:

• Monitoring underground work environments is crucial for industrial safety.
• Detecting safety helmet wearing faces challenges like complex backgrounds, low light, and small target detection.
• Existing methods struggle with multiscale feature fusion, foreground localization, and dynamic context modeling.

Purpose of the Study:

• To propose a robust object detection model for safety helmet detection in challenging underground environments.
• To address limitations in existing methods concerning multiscale feature fusion and localization accuracy.
• To enhance the detection of small targets and improve overall model performance in industrial safety applications.

Main Methods:

• Developed the YOLOv11-SRA model, integrating SAConv, RCM, and ASFF modules.
• SAConv dynamically adjusts dilation rates for multiscale contextual information and small target detection.
• RCM refines foreground regions using rectangular self-calibrated attention for improved boundary localization.
• ASFF fuses multiscale features via adaptive spatial weighting to mitigate feature conflicts.

Main Results:

• The YOLOv11-SRA model achieved a mean average precision (mAP50) of 84.2% on the CUMT-HelmeT dataset.
• The model demonstrated a recall of 79.9%, significantly outperforming mainstream object detection models.
• Validated effectiveness in complex underground safety helmet detection scenarios.

Conclusions:

• The proposed YOLOv11-SRA model effectively addresses challenges in safety helmet detection.
• The integrated optimization strategy enhances robustness, localization accuracy, and multiscale feature fusion.
• This model offers a significant advancement for improving safety monitoring in industrial environments.