Vision-based work posture assessment: monocular pose estimation and fuzzy REBA methods
开发了一种基于单目摄像头的无接触姿势风险评估系统，通过改进的姿态估计模型和模糊逻辑REBA方法，在工业环境中实现实时、准确的工效学监测，降低肌肉骨骼疾病风险。
With the implementation of China's National Occupational Disease Prevention Plan (2021-2025), workers' health has gained growing attention. To reduce incidence of work-related musculoskeletal disorders (MSDs), this study developed a contactless, vision-based system for posture risk assessment. An MMPose-based monocular pose estimation model was designed to extract skeletal landmarks for joint-angle computation, and it demonstrated the accuracy superior to that of Kinect V2 in challenging industrial environments. Experimental validation demonstrated that the proposed model achieved an average joint-angle root mean square error (RMSE) of 5.7° (lower than the 10.7° of Kinect V2), with 91% of joints exhibiting errors below 8°. Furthermore, the enhanced fuzzy-logic REBAPRO reduced false risk transitions by 38% and increased expert agreement from 67% to 92%, with the system achieved an inference latency of 6.9 ms/frame to support real-time industrial monitoring. Deployment in an automotive manufacturing plant confirmed its potential for real-time ergonomic monitoring and proactive risk mitigation.