Ergonomic investigations on novel dynamic postural estimator using blaze pose and transfer learning

Area of Science:

• Ergonomics
• Computer Vision
• Biomechanics

Background:

• Dynamic work postures are crucial for assessing musculoskeletal risk but lack comprehensive evaluation methods.
• Existing methods for postural analysis are often time-consuming and may not capture dynamic movements effectively.

Purpose of the Study:

• To develop and validate a novel computer-based model for dynamic postural evaluation.
• To leverage BlazePose and transfer learning for accurate human pose estimation in ergonomic assessments.
• To provide a tool for rapid online and offline analysis of dynamic work postures using the RULA (Rapid Upper Limb Assessment) framework.

Main Methods:

• Development of a camera-based, three-dimensional (3D) dynamic human pose estimation model using BlazePose.
• Utilizing a dataset of 50,000 action-level images for model training.
• Application of Deep Neural Network (DNN) and Transfer Learning (TL) approaches.
• Integration with the RULA framework for ergonomic assessment.

Main Results:

• The developed model achieved a promising accuracy of 94.12% for dynamic postural assessment.
• High accuracy, precision, and recall were demonstrated for each output prediction class.
• The model effectively analyzed ergonomics of dynamic postures, considering factors like muscle loading and foot support.

Conclusions:

• A novel and accurate dynamic postural estimator using BlazePose and transfer learning has been successfully developed.
• The computer-based model offers a viable solution for efficient and precise ergonomic evaluation of dynamic work postures.
• This approach addresses the need for detailed investigation into dynamic work postures, enhancing occupational safety assessments.