Based on the linked source; automatically prepared and checked against the original report.
Study design
The study presents a vision-based framework for assessing ergonomic risk during cabinet interaction. Its dataset included 30 participants: 14 experienced wheelchair users and 16 trained simulation participants. Researchers recorded approximately five hours of RGB video across five cabinet-operation scenarios, then selected 10,000 representative images from about 150,000 original frames for annotation and model development.
Technical approach
The framework combines YOLOv11 for human detection, MHFormer for monocular three-dimensional pose reconstruction, and a fuzzy-logic-enhanced RULA model for continuous risk quantification. According to the research summary, the system converts video-derived motion signals into temporally continuous kinematic representations and ergonomic risk scores, supporting non-contact assessment in assistive-living environments.
Reported findings
The proposed method produced an average joint-angle estimation RMSE of 7.5 degrees, compared with 18.6 degrees for a Kinect v2-based baseline. It achieved 84% risk-classification accuracy and a Cohen’s kappa of 0.66 in benchmarking. The study also reported higher and more sustained risk during low revolving-door and low-drawer operations than during sliding-door interaction, while presenting the results as evidence for monitoring and assessment rather than treatment guidance.
