ViDMASK dataset for face mask detection with social distance measurement

Area of Science:

• Computer Science
• Artificial Intelligence
• Public Health

Background:

• The COVID-19 pandemic highlighted the need for automated monitoring of public health measures.
• Existing video surveillance systems can be enhanced with AI for real-time adherence tracking.

Purpose of the Study:

• To develop and evaluate an AI-based system for detecting face mask adherence and measuring social distancing.
• To introduce and validate a new large-scale dataset, ViDMASK, for training and testing these models.

Main Methods:

• Deep learning models, including Mask RCNN, YOLOv4, YOLOv5, and YOLOR, were trained for object detection and recognition.
• Models were evaluated on the MOXA dataset and the newly proposed ViDMASK dataset.
• Post-detection, inter-personal distances were calculated to assess social distancing compliance.

Main Results:

• The YOLOR model achieved a mean average precision of 92.4% for face mask detection on existing datasets.
• Models performed well on the MOXA dataset but showed reduced accuracy on the more complex ViDMASK dataset.
• The ViDMASK dataset presents challenges due to diverse subject characteristics, poses, environments, and image quality.

Conclusions:

• AI models demonstrate potential for monitoring face mask adherence and social distancing in public spaces.
• The ViDMASK dataset offers a valuable resource for advancing research in real-world public health surveillance.
• Further model refinement is needed to address the complexities of diverse real-world scenarios.