Lightweight Vision Transformer for Frame-Level Ergonomic Posture Classification in Industrial Workflows.

Area of Science:

• Industrial Ergonomics
• Computer Vision
• Occupational Safety

Background:

• Work-related musculoskeletal disorders (WMSDs) are a significant concern in industrial settings.
• These disorders often result from prolonged non-neutral postures and repetitive tasks.
• Current assessment methods can be complex and intrusive.

Purpose of the Study:

• To develop a vision-based framework for real-time, frame-level ergonomic risk classification.
• To utilize a lightweight Vision Transformer (ViT) for posture assessment directly from RGB images.
• To enable unobtrusive and scalable monitoring of ergonomic risks in manufacturing environments.

Main Methods:

• A Vision Transformer (ViT) model was employed for direct analysis of raw RGB images.
• The system simultaneously classifies eight anatomical regions for multi-label posture assessment.
• Training utilized a multimodal dataset with synchronized RGB video and motion capture data, with labels derived from RULA scores.

Main Results:

• The ViT model achieved state-of-the-art performance, with F1-scores > 0.99 and AUC values > 0.996 across all anatomical regions.
• The system demonstrated high accuracy and generalizability on simulated industrial tasks, including those with occlusion and posture variability.
• Compared to CNN-based systems, the ViT model reduced complexity and enabled real-time inference on edge devices.

Conclusions:

• The proposed vision-based framework offers an effective solution for real-time ergonomic risk classification.
• The ViT model provides a lightweight, accurate, and generalizable approach to posture assessment.
• This technology holds significant potential for improving worker safety and reducing WMSDs in manufacturing.