Automatic estimation of Hand Activity Level from upper-limb trajectories: a probabilistic regression framework

Accurate measurement of Hand Activity Level (HAL) is crucial for evaluating musculoskeletal injury risk in repetitive hand-intensive work. Manual HAL assessments are often subjective and impractical for large-scale or continuous monitoring. This study presents a probabilistic regression framework that leverages video-based upper-limb pose trajectories to automatically estimate HAL scores while providing associated confidence measures. By enabling ergonomic risk assessment with quantified uncertainty, the proposed method delivers objective and reliable HAL predictions. Experimental results demonstrate strong in-domain performance (Root Mean Square Error [RMSE] = 0.24, Mean Absolute Error [MAE] = 0.17) and robust cross-domain generalisation (RMSE = 0.74, MAE = 0.54), highlighting both the accuracy and transferability of the framework.