The accuracy of several pose estimation methods for 3D joint centre localisation.

Area of Science:

• Biomechanics
• Human Movement Analysis
• Computer Vision

Background:

• Traditional motion capture is lab-bound and resource-intensive.
• Markerless pose estimation offers potential for 'in-the-wild' human movement studies.
• The accuracy of markerless systems requires thorough evaluation.

Purpose of the Study:

• To evaluate the accuracy of markerless deep-learning pose estimation algorithms.
• To compare 3D joint center locations derived from markerless methods against marker-based motion capture.
• To identify discrepancies in joint center accuracy across different activities.

Main Methods:

• Utilized OpenPose, AlphaPose, and DeepLabCut for 3D joint center estimation.
• Collected synchronized marker-based motion capture and multi-camera high-speed (200 Hz) imaging data.
• Applied algorithms to 2D image data and reconstructed 3D joint locations for walking, running, and jumping.

Main Results:

• Systematic differences of 30-50 mm observed at the hip and knee, likely due to training data mislabeling.
• Smaller differences (1-15 mm) noted at the ankle, varying by activity.
• Markerless systems showed activity-dependent accuracy compared to marker-based systems.

Conclusions:

• Markerless motion capture is a promising technology for liberating movement research from laboratory settings.
• Current 3D joint center accuracy from markerless systems is not consistently comparable to marker-based methods.
• Further refinement of algorithms and training data is needed for widespread adoption.