Computer Vision–Based monocular 3D human pose estimation and ergonomic risk assessment: The impact of model variants and camera views on kinematic accuracy and ergonomic assessment validity

Work-related musculoskeletal disorders (WMSDs) remain highly prevalent and are a leading cause of occupational disability, underscoring the need for scalable and reliable ergonomic risk assessment methods. Although computer vision has shown great promise in automating human pose estimation and ergonomic risk evaluation, systematic evidence is still required to clarify how model variant, camera view, and occlusion affect both kinematic accuracy and ergonomic assessment outcomes. This study systematically compared two widely used open-source monocular 3D human pose estimation models, MediaPipe and MMPose , against an optical motion capture reference system across a series of static and dynamic tasks recorded from three camera views (front, 45°, and side). Results showed that MediaPipe achieved higher overall kinematic accuracy in estimating joint positions and angles, as well as stronger agreement with ergonomic risk assessment scores (REBA and RULA) than MMPose . MediaPipe achieved optimal performance at the 45° camera view for both joint estimation and ergonomic assessment, whereas MMPose performed best from the side view for pose estimation and the front view for ergonomic assessment. These findings clarify how pose-estimation accuracy translates into ergonomic assessment validity and provide practical guidance for applying computer vision–based methods to achieve accessible, scalable, and reliable ergonomic risk assessment in workplace settings.