Research on application of helmet wearing detection improved by YOLOv4 algorithm

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• The YOLOv4 algorithm suffers from a high parameter count and suboptimal small target detection.
• Effective helmet detection is crucial for workplace safety and regulatory compliance.

Purpose of the Study:

• To develop an improved YOLOv4-based model for helmet detection with enhanced accuracy and efficiency.
• To address the limitations of the original YOLOv4 algorithm concerning parameter size and small object recognition.

Main Methods:

• Incorporated multi-scale prediction and an enhanced PANet structure to boost small target detection accuracy.
• Replaced standard convolutions with depth-separable convolutions to significantly reduce model parameters.
• Utilized k-means clustering for prior box optimization.
• Evaluated the model on a custom helmet dataset (helmet_dataset).

Main Results:

• The improved YOLOv4 model achieved a mean Average Precision (mAP) of 93.05%, a 0.49% increase over the original YOLOv4.
• Model parameters were reduced by approximately 58% to about 105 MB.
• The model demonstrated a detection speed of 35 Frames Per Second (FPS).
• Outperformed the Faster R-CNN algorithm in detection speed, accuracy, and parameter count.

Conclusions:

• The proposed improved YOLOv4 algorithm offers a more efficient and accurate solution for helmet detection.
• The model's reduced parameter count and improved performance make it suitable for real-world deployment in various scenarios.
• This advancement contributes to enhanced safety monitoring through effective helmet usage detection.