A Deep Learning Model for Markerless Pose Estimation Based on Keypoint Augmentation: What Factors Influence Errors in...

Area of Science:

• Biomechanics
• Computer Vision
• Deep Learning

Background:

• Traditional marker-based motion capture in biomechanics presents several inconveniences.
• Markerless methods, like pose estimation networks, offer an alternative but often lack complete anatomical data for joint angle calculations.

Purpose of the Study:

• To develop and evaluate deep learning-based marker augmentation models for accurate markerless motion capture.
• To compare the performance of different complexity marker augmentation models against a photogrammetry system.

Main Methods:

• Three marker augmentation models of varying complexity were developed.
• Models were trained and tested using data compared against a marker-based photogrammetry system.
• Statistical analysis was performed to identify factors influencing position and joint angle errors.

Main Results:

• The proposed Transformer model achieved position errors under 1.5 cm for anatomical landmarks and 4.4 degrees for joint angles across seven movements.
• Anthropometric data did not significantly influence errors.
• Anatomical landmarks and movement type affected position errors, while model, rotation axis, and movement influenced joint angle errors.

Conclusions:

• Deep learning marker augmentation models can significantly improve the accuracy of markerless motion capture in biomechanics.
• The Transformer model demonstrates superior performance, providing comprehensive anatomical data for detailed movement analysis.