A biomechanical-constrained temporal learning framework for lightweight skeleton-based exercise recognition.

Area of Science:

• Biomechanics
• Deep Learning
• Computer Vision
• Human Motion Analysis

Background:

• Skeletal pose recognition is crucial for fitness technology, rehabilitation, and sports analytics.
• Current methods often overlook fundamental biomechanical principles of human movement.
• Integrating biomechanics can enhance the accuracy and robustness of exercise recognition systems.

Purpose of the Study:

• To present the Biomechanical-Aware Temporal Learning (BATL) framework for exercise recognition.
• To incorporate human kinematic constraints and deep temporal models into skeletal pose analysis.
• To improve the accuracy and efficiency of recognizing exercises by considering biomechanical limitations.

Main Methods:

• Developed the Biomechanical-Aware Temporal Learning (BATL) framework, integrating kinematic constraints (joint angle consistency, velocity smoothness, bone length stability).
• Employed an unsupervised phase discovery module to automatically identify temporal divisions (preparation, execution, recovery) without manual annotation.
• Validated the framework on diverse exercise datasets (squats, push-ups, bicep curls, shoulder-presses).

Main Results:

• Achieved an estimated test accuracy of 93.33% ± 0.94% via 5-fold cross-validation.
• Outperformed baseline techniques by 6.22%-14.44%, with biomechanical constraints improving accuracy by 7.78%.
• Demonstrated efficient inference (5.2 ms per 30-frames sequence in BatL-only mode) with a compact model size (3.57M parameters, 13.6 MB memory).

Conclusions:

• Biomechanical knowledge integrated with deep learning offers significant prospects for activity recognition systems.
• The BATL framework effectively enhances movement quality and temporal analysis in exercise recognition.
• Domain knowledge, even as inductive bias, can be more effective than solely relying on complex models.