COVID-19 risk reduce based YOLOv4-P6-FaceMask detector and DeepSORT tracker

Area of Science:

• Computer Vision
• Artificial Intelligence
• Public Health Technology

Background:

• Effective mask-wearing is crucial for public health, yet automated detection and tracking systems are underdeveloped.
• Existing image processing research lacks robust solutions for real-time facemask status identification.

Purpose of the Study:

• To develop a high-performance, two-stage facemask detector and tracker using a monocular camera and deep learning.
• To automate the detection and tracking of facemask usage in video sequences.

Main Methods:

• Utilized a novel dataset of 18,000 images with over 30,000 annotations for 'face masked', 'incorrectly masked', and 'no masked' classes.
• Employed the Scaled-You Only Look Once (Scaled-YOLOv4) model for detection (YOLOv4-P6-FaceMask) and Simple Online and Real-time Tracking with a deep association metric (DeepSORT) for tracking.
• Implemented DeepSORT for efficient face tracking and database creation of individuals without masks.

Main Results:

• The YOLOv4-P6-FaceMask detector achieved 93% mean average precision and 92% mean average recall.
• The system demonstrated real-time processing speeds of 35 frames per second on a single Tesla-T4 GPU.
• Performance was validated against state-of-the-art facemask detection and tracking models.

Conclusions:

• The proposed deep learning framework offers a highly accurate and efficient solution for automated facemask detection and tracking.
• This technology can significantly enhance public health surveillance and compliance monitoring in real-world scenarios.