Assessment of monocular human pose estimation models for clinical movement analysis

Area of Science:

• Computer Vision
• Biomechanical Analysis
• Machine Learning

Background:

• Markerless human pose estimation (HPE) is emerging as a cost-effective alternative to marker-based motion capture.
• While offering benefits like low hardware needs and faster setup, markerless HPE often lags in accuracy and speed.
• This performance gap can limit the application scope of markerless techniques.

Purpose of the Study:

• To comprehensively assess the accuracy, precision, and inference speed of 11 open-source monocular markerless HPE methods.
• To compare these estimators against gold-standard marker-based optical motion capture.
• To evaluate their suitability for analyzing physical activities.

Main Methods:

• Creation of Physio2.2M, a dataset with 2.2 million RGB frames from 25 participants performing exercises.
• Ground truth data obtained using a marker-based optical motion capture system.
• Evaluation of 11 monocular markerless HPE algorithms on the Physio2.2M dataset.

Main Results:

• Mean per joint position error ranged from 72-122 mm (2D) and 146-249 mm (3D).
• Knee and elbow flexion angles showed mean absolute errors of [Formula: see text] (2D) and [Formula: see text] (3D).
• Inference speeds varied: direct estimators (25-200 FPS), 2D-to-3D lifting methods (117-9341 FPS).

Conclusions:

• Performance (accuracy, precision, speed) varies significantly across markerless HPE methods and image dimensions.
• Some 2D estimators approach visual assessment accuracy.
• Current markerless HPE methods show promise but have limitations for precise biomechanical analysis in physical activities.