A Skeleton-Based Rehabilitation Exercise Assessment System With Rotation Invariance

Area of Science:

• Biomedical Engineering
• Computer Vision
• Rehabilitation Science

Background:

• Automated exercise assessment is crucial for patients undergoing rehabilitation, requiring professional guidance.
• Skeleton-based models are popular for exercise assessment due to ease of implementation but face challenges with orientation sensitivity and lack of feedback.
• Existing methods struggle with variations in human skeleton orientation and do not provide specific feedback on incorrect movements.

Purpose of the Study:

• To develop a novel rotation-invariant descriptor for skeleton-based exercise assessment.
• To address the sensitivity of current models to human skeleton orientation.
• To introduce a visualization method for providing corrective feedback to users during exercise.

Main Methods:

• Proposed a novel rotation-invariant descriptor: the dot product matrix of the human skeleton.
• Developed a visualization technique using Gradient-Weighted Class Activation Mapping (Grad-CAM).
• Introduced a quantitative metric, Overlap Ratio (OvR), to evaluate visualization quality.
• Conducted experiments on public datasets and a self-generated push-up dataset.

Main Results:

• The proposed rotation-invariant descriptor demonstrated absolute robustness to orientation, even with significant angle perturbations.
• The method achieved superior accuracy and Overlap Ratio (OvR) compared to previous works in most cases.
• Visualization results effectively identified incorrect movements, offering informative feedback for users.

Conclusions:

• The novel rotation-invariant descriptor enhances the stability and accuracy of skeleton-based exercise assessment models.
• The Grad-CAM based visualization provides valuable, actionable feedback for user movement correction in rehabilitation.
• This approach offers a robust and informative solution for automated exercise assessment in clinical and home settings.