HDS-YOLOv5: An improved safety harness hook detection algorithm based on YOLOv5s.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Safety

Background:

• Improper safety harness hook usage is a significant safety hazard in power maintenance.
• Traditional machine vision methods lack accuracy and real-time effectiveness for hook status detection.

Purpose of the Study:

• To develop a novel deep learning network, HDS-YOLOv5, for accurate and real-time detection of safety harness hook status.
• To reduce safety incidents in complex power maintenance environments.

Main Methods:

• Modified YOLOv5s architecture incorporating a HOOK-SPPF feature extraction module.
• Implemented a decoupled head module to improve classification and regression.
• Utilized Scylla Intersection over Union (SIoU) to optimize the loss function.

Main Results:

• The HDS-YOLOv5 algorithm achieved a 3% increase in mean Average Precision (mAP@0.5), reaching 91.2%.
• Demonstrated a detection rate of 24.0 frames per second (FPS).
• Showcased superior performance compared to existing models.

Conclusions:

• HDS-YOLOv5 offers enhanced feature extraction and recognition accuracy for safety harness hooks.
• The improved model effectively addresses limitations of traditional methods in complex environments.
• The algorithm contributes to reducing safety incidents through reliable, real-time monitoring.