Abstract
While learnable adjacency matrices have been explored to enhance the flexibility of Spatio-Temporal Graph Convolutional Networks (ST-GCNs) for action recognition, their application in wearable sensor-based systems often overlooks a critical constraint: the need to maintain biomechanically plausible connections while adapting to non-standard sensor placements. To address this, we propose a dynamic topology-adaptive ST-GCN framework that strategically initializes the learnable adjacency matrix with a human skeleton prior. This ensures that the initial graph structure is physiologically meaningful. Subsequently, the model refines this topology through end-to-end training, incorporating L2 regularization and periodic Top-K sparsification to prevent overfitting and maintain a sparse, interpretable structure. This approach allows the model to dynamically correct for structural deviations caused by variations in sensor positioning across individuals, without deviating from realistic body kinematics. Evaluated on data from eight IMU sensors, our method achieves an average recognition accuracy of 94.1 ± 0.6% in cross-user scenarios and 91.5 ± 0.5% in cross-device tests, demonstrating superior robustness for sports posture recognition under non-standardized deployment conditions.