Load Asymmetry Angle Estimation Using Multiple view Videos

Area of Science:

• Ergonomics and Biomechanics
• Computer Vision and Machine Learning
• Occupational Safety and Health

Background:

• Repetitive lifting tasks pose risks for lower back injuries.
• The revised NIOSH lifting equation (RNLE) uses load asymmetry angle to assess these risks.
• Accurate estimation of the asymmetry angle is crucial for workplace safety.

Purpose of the Study:

• To develop a robust computer vision approach for estimating the load asymmetry angle.
• To improve the accuracy and reliability of angle estimation in lifting assessments.
• To mitigate challenges posed by occlusions in video data.

Main Methods:

• Utilized OpenPose for 2D skeletal joint estimation from synchronized videos.
• Developed a computer vision correspondence and depth estimation method for 3D joint coordinates.
• Implemented a novel method to flag unreliable angle estimates based on joint confidence levels.

Main Results:

• Reduced the standard deviation of angle estimation error from 10.13° to 4.99° in a dataset of 360 lifting instances.
• Identified an optimal threshold to balance error reduction and flagging of unreliable estimates.
• Successfully flagged 34% of estimates requiring further validation, enhancing overall reliability.

Conclusions:

• The developed computer vision method significantly enhances the accuracy of load asymmetry angle estimation.
• The flagging mechanism improves the robustness of the lifting assessment tool.
• This approach offers a promising solution for objective and reliable ergonomic risk assessment in occupational settings.