3D human pose keypoints and corresponding joint angle calculation for vision-based WMSD risk assessments.

Area of Science:

• Biomechanical Engineering
• Computer Vision
• Occupational Health

Background:

• Vision-based pose estimation offers an accessible method for assessing work-related musculoskeletal disorder (WMSD) risks.
• Current models lack 3D data for precise joint angle calculations crucial for WMSD assessment.

Purpose of the Study:

• To develop a specialized keypoint set and methodology for accurate 3D joint angle calculation using vision-based pose estimation.
• To enhance WMSD risk assessment by improving the accuracy of biomechanical data derived from videos.

Main Methods:

• Defined a 66-keypoint set optimized for joint angle calculations and visual feature extraction.
• Collected a large dataset (6.7 million frames) of manual material handling and assembly tasks.
• Trained a baseline pose estimation model for calculating joint angles from the specialized keypoints.

Main Results:

• Enabled calculation of 22 angles across 6 body joints for WMSD risk assessment.
• Achieved a mean absolute angle error of 2.4° with a baseline pose estimation model.
• Demonstrated the utility of the specialized keypoint set for accurate joint angle calculations.

Conclusions:

• The specialized keypoint set and methodology are effective for calculating joint angles from videos.
• This approach improves the feasibility of vision-based WMSD risk assessment.
• Further development can enhance the accuracy and application of this method.