A Movement Description Language for Functional Training Exercise Analysis.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Exercise Science

Background:

• Traditional exercise recognition systems struggle with complex functional movements.
• Need for interpretable and accurate methods for analyzing multi-joint exercises.

Purpose of the Study:

• Introduce a Movement Description Language (MDL) for formal representation and analysis of functional exercises.
• Develop a system for interpretable automatic repetition counting and phase detection using camera-based pose estimation.

Main Methods:

• Modeled exercises as finite-state machines using pose-derived angle proxy transitions.
• Utilized MediaPipe for landmark extraction from monocular video.
• Validated on 1513 videos of 12 functional exercises with real-world variations.

Main Results:

• Achieved 97.2% overall accuracy in automated repetition counting against ground truth.
• Demonstrated high per-exercise accuracy (mean 98.8%) across diverse movements.
• Showcased reliable phase detection despite variations in speed, perspective, and environment.

Conclusions:

• The MDL framework offers a transparent, extensible, and efficient solution for functional exercise analysis.
• Enables interpretable automatic repetition counting and phase detection, serving as an alternative to black-box methods.
• Potential for scalable deployment in fitness training, monitoring, and movement analysis applications.