Multi-Scale Safety Helmet Detection Based on RSSE-YOLOv3

Area of Science:

• Computer Vision
• Artificial Intelligence
• Occupational Safety

Background:

• Real-time detection of safety helmet compliance in construction is challenging due to occlusions and small target sizes, leading to low accuracy and missed detections.
• Existing object detection algorithms like YOLOv3 struggle with dense and small targets, necessitating algorithmic improvements for enhanced safety monitoring.

Purpose of the Study:

• To develop a novel detection algorithm based on YOLOv3 to improve the accuracy and efficiency of detecting construction workers wearing safety helmets.
• To address the limitations of low detection accuracy and missed detections in current safety helmet monitoring systems.

Main Methods:

• The original YOLOv3 network's Res8 module was replaced with a parallel RepVGG Skip Squeeze Excitation (RSSE) module to increase network width and speed.
• Residual module Res2 replaced CBL×5 modules to prevent gradient disappearance and enhance feature reuse.
• Input image resolution was increased, and four-scale feature prediction was adopted for improved small target detection.
• Complete Joint Crossover (CIOU) was integrated to enhance the loss function and localization accuracy.

Main Results:

• The improved YOLOv3 algorithm demonstrated a 3.9% increase in precision (P).
• Recall (R) improved by 5.2%, and mean average precision (mAP) increased by 4.7% compared to the original YOLOv3.
• The enhanced algorithm significantly improved the overall performance in detecting safety helmets.

Conclusions:

• The proposed YOLOv3-based algorithm effectively enhances the detection of safety helmets in construction environments.
• The integration of RSSE modules, Res2 modules, and CIOU loss function leads to significant improvements in detection speed and accuracy.
• This advancement contributes to more reliable real-time safety monitoring systems in the construction industry.