A Deep Neural Network-based method for estimation of 3D lifting motions

Area of Science:

• Biomechanics
• Computer Vision
• Machine Learning

Background:

• Marker-based motion capture systems have limitations including time-consuming setup, movement obstruction, and strict environmental requirements.
• Accurate 3D pose estimation is crucial for analyzing biomechanics during tasks like lifting.
• Developing markerless systems is essential for practical, on-site biomechanical assessments.

Purpose of the Study:

• To develop and validate a Deep Neural Network (DNN) for markerless 3D pose estimation during lifting.
• To overcome the constraints associated with traditional marker-based motion capture techniques.
• To provide an accurate and accessible tool for on-site biomechanical analysis.

Main Methods:

• A two-stage cascaded Deep Neural Network (DNN) was designed, augmenting an Hourglass network for 2D pose estimation with a 3D pose generator.
• The DNN utilized images from two camcorders to synthesize information from multiple views for 3D pose prediction.
• Twelve healthy adults performed nine lifting tasks under varying heights and asymmetry angles, with data validated against a marker-based motion capture system.

Main Results:

• The proposed DNN achieved an average 3D pose error of 14.72 ± 2.96 mm compared to the marker-based system.
• Lifting conditions, such as 60° asymmetry and shoulder-height lifting, resulted in higher pose estimation errors.
• The method demonstrated high accuracy and robustness across different lifting scenarios.

Conclusions:

• The developed DNN-based method provides accurate 3D pose estimation for lifting tasks without marker-based system limitations.
• This markerless approach offers a practical solution for on-site biomechanical analysis.
• The findings support the potential application of this technology in ergonomic assessments and injury prevention.