Joint Angle Estimation from RGB-D Fusion for 3D Automated Ergonomic Assessment

Abstract

Musculoskeletal disorders (MSDs) remain a critical concern in industrial settings. Although observational methods such as REBA and RULA are widely used for postural risk assessment, they rely on expert judgment and are time-consuming. We present an automated method that combines 2D keypoint detection via OpenPose with depth information from an RGB-D camera to estimate 3D joint angles relevant to ergonomic scoring. A polynomial regression filter and a neutral-posture calibration phase enhance robustness and reduce noise. Validation experiments on representative industrial tasks demonstrate that 3D angle estimates significantly outperform 2D estimates, particularly during dynamic movements, as confirmed by VICON reference data and expert annotations. Principal Component Analysis indicates that 3D data provide a richer and less redundant representation of posture. Moreover, REBA 2D scores consistently underestimate ergonomic risk compared to REBA 3D scores, especially in high-risk scenarios. This approach enables more precise detection of unsafe postures and can be integrated into automated ergonomic assessment workflows to help prevent MSDs in real-world environments.

Dwayne Lauret ¹, Abderraouf Benali ¹ Halim Djerroud ¹,

¹Université Paris-Saclay, UVSQ, LISV, 78140, Vélizy-Villacoublay, France.