Automated Physical Distance Estimation and Crowd Monitoring Through Surveillance Video.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Public Health

Background:

• COVID-19 transmission is reduced by physical distancing, recommended by WHO.
• Maintaining the 2-m distance is challenging in crowded public spaces like malls.
• Active monitoring systems are needed to enforce social distancing and slow disease spread.

Purpose of the Study:

• To develop a deep learning system for automatic physical distance detection using security camera footage.
• To enhance object detection and tracking for accurate social distancing monitoring.
• To identify individuals violating physical distancing rules in real-time.

Main Methods:

• Utilized TH-YOLOv5 for object detection and classification, incorporating Transformer Heads (TH) and CBAM for improved accuracy.
• Employed Deepsort for tracking detected individuals using bounding boxes.
• Implemented pairwise L2 vectorized normalization for 3D feature space tracking and violation index calculation.

Main Results:

• The system achieved a weighted mAP score of 89.5% for accurate detection and classification.
• Achieved an FPS score of 29, indicating efficient real-time processing.
• Demonstrated computational comparability with existing methods.

Conclusions:

• The proposed deep learning system effectively monitors physical distancing in public areas.
• The system aids in enforcing COVID-19 safety precautions and reducing virus transmission.
• This technology offers a viable solution for active public health monitoring.