Automated Assessment of PPE Compliance with Fast Lightweight Deep Learning Based Computer Vision.

Area of Science:

• Computer Vision
• Machine Learning
• Occupational Health and Safety

Background:

• Workplace injuries and occupational diseases can be reduced by over 37% through appropriate Personal Protective Equipment (PPE) usage.
• Previous research on PPE compliance using models like RepVGG and ResNet50 reported limited accuracy (around 80%) and faced challenges with large model sizes and high inference times.

Purpose of the Study:

• To develop a novel, lightweight computer vision framework for automated PPE compliance verification before worksite entry.
• To achieve over 90% classification accuracy with faster inference times than standard Vision Transformers, enabling deployment on consumer-grade hardware.
• To identify incorrect PPE usage by analyzing the spatial relationship between workers and their equipment.

Main Methods:

• A novel framework incorporating a state space model layer for linear computation and a spatial attention module for global dependency capture.
• Utilized positional encoding within the image processing pipeline.
• Evaluated on two public datasets (PPE-CLS and PPE-BQZEL) focusing on goggles and N95 mask compliance in manufacturing settings.

Main Results:

• Achieved accuracies of 0.91 and 0.94 on the two datasets, outperforming a standard Vision Transformer (ViT) which reported 0.88 and 0.90.
• The model operated up to 5 times faster than the benchmark ViT.
• The framework boasts a minimal size (2.8M parameters) and linear inference scaling for real-time performance.

Conclusions:

• Presents a fast, memory-efficient solution for automated safety supervision in occupational settings.
• Advances the application of artificial intelligence and machine learning in improving occupational health and safety compliance.