A new YOLO-based method for social distancing from real-time videos

Area of Science:

• Computer Vision
• Deep Learning
• Public Health Technology

Background:

• COVID-19 transmission highlights the need for effective physical distancing measures.
• Maintaining at least one meter distance between individuals in indoor settings is a key precaution.
• Current monitoring methods may lack real-time capabilities for ensuring compliance.

Purpose of the Study:

• To propose and evaluate a real-time method for monitoring physical distancing compliance in indoor environments.
• To leverage computer vision and deep learning for automated detection and distance estimation.
• To assess the performance of different YOLO models for this specific application.

Main Methods:

• Utilized the YOLO (You Only Look Once) object detection model, a convolutional neural network.
• Pre-trained the YOLO model on the Microsoft COCO dataset for person detection.
• Implemented real-time analysis to estimate physical distances between detected individuals.
• Evaluated performance using accuracy, frames per second (FPS), and mean average precision (mAP).

Main Results:

• YOLO v3 demonstrated high accuracy (87.07%) and mean average precision (mAP) (89.91%).
• YOLO v5s achieved the highest frame rate (FPS) of 18.71, indicating real-time processing efficiency.
• The study successfully validated the proposed deep learning approach for distance monitoring.

Conclusions:

• The proposed computer vision and deep learning method is effective for real-time monitoring of physical distancing.
• YOLO models offer a viable solution for automated compliance checks in public spaces.
• Findings provide a foundation for applying similar technologies in future public health surveillance scenarios.