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

Area of Science:

• Ergonomics and Occupational Health
• Computer Vision and Machine Learning
• Biomechanical Engineering

Background:

• Accurate Hand Activity Level (HAL) measurement is vital for assessing musculoskeletal injury risk in repetitive tasks.
• Current manual HAL assessments are subjective and not scalable for continuous monitoring.
• Objective and automated methods are needed for reliable ergonomic risk evaluation.

Purpose of the Study:

• To develop and validate a probabilistic regression framework for automatic HAL score estimation using video data.
• To provide confidence measures alongside HAL predictions for quantified uncertainty in ergonomic risk assessment.
• To enable objective, reliable, and scalable monitoring of hand activity in occupational settings.

Main Methods:

• Utilized video-based upper-limb pose trajectories as input features.
• Developed a probabilistic regression framework for HAL score estimation.
• Incorporated confidence measures to quantify prediction uncertainty.

Main Results:

• Achieved strong in-domain performance with Root Mean Square Error (RMSE) = 0.24 and Mean Absolute Error (MAE) = 0.17.
• Demonstrated robust cross-domain generalisation with RMSE = 0.74 and MAE = 0.54.
• The framework provides accurate and transferable HAL predictions.

Conclusions:

• The proposed video-based probabilistic regression framework enables objective and reliable automatic estimation of Hand Activity Level (HAL).
• Quantified uncertainty through confidence measures enhances ergonomic risk assessment.
• The method shows significant potential for large-scale and continuous monitoring of hand-intensive work, improving occupational safety.