Wearable sensor data-driven sports posture recognition using the ST-GCN spatio-temporal graph convolutional network.

Area of Science:

• Biomechanical Engineering
• Computer Science
• Machine Learning

Background:

• Spatio-Temporal Graph Convolutional Networks (ST-GCNs) offer flexibility in action recognition.
• Learnable adjacency matrices in ST-GCNs face challenges with non-standard sensor placements in wearable systems.
• Maintaining biomechanically plausible connections is crucial for accurate analysis.

Purpose of the Study:

• To propose a dynamic topology-adaptive ST-GCN framework for robust action recognition using wearable sensors.
• To ensure physiologically meaningful graph structures by initializing with a human skeleton prior.
• To dynamically adapt the topology for variations in sensor placement without compromising kinematic realism.

Main Methods:

• Developed a dynamic topology-adaptive ST-GCN framework.
• Initialized the learnable adjacency matrix with a human skeleton prior.
• Employed end-to-end training with L2 regularization and Top-K sparsification for structural refinement and interpretability.

Main Results:

• Achieved 94.1% accuracy in cross-user scenarios with eight IMU sensors.
• Attained 91.5% accuracy in cross-device tests.
• Demonstrated superior robustness for sports posture recognition under non-standardized deployment conditions.

Conclusions:

• The proposed framework effectively addresses the challenge of non-standard sensor placements in wearable-based action recognition.
• Dynamic topology adaptation, guided by biomechanical priors, enhances model robustness and accuracy.
• The method ensures physiologically meaningful and interpretable graph structures for reliable human activity analysis.