Time-warping Augmentation and ST-GCN for Physical Rehabilitation Exercise Assessment

This paper presents an effective and scalable framework for evaluating physical rehabilitation exercises, with a focus on enhancing prediction accuracy and providing clinically meaningful feedback to support patient recovery. The proposed method comprises three key components. First, TimeWarping Augmentation is introduced to simulate variations in movement speed, improving the model’s robustness to diverse exercise tempos. Second, a feature extraction pipeline is designed, integrating spatial enhancements via channel and joint attention mechanisms with Multi-Scale Temporal
Convolutional Blocks to capture temporal dynamics at multiple resolutions. Third, a sequence modeling module is employed,incorporating joint-wise feature aggregation, positional encoding, bidirectional LSTM, and temporal attention pooling to extract fine-grained temporal patterns. The framework is evaluated on the KIMORE datasets, achieving competitive performance. Notably, it achieves a 21.62% reduction in Mean Absolute Deviation (MAD) compared to existing methods. These results underscore the potential of the proposed system for realworld telerehabilitation applications, enabling reliable and automated assessment in unsupervised environments.