Spatio-Temporal Transformer Model for Skeleton-based Rehabilitation Exercises Assessment

Physical rehabilitation training is a critical process that enables patients with chronic illnesses and elderly individuals to recover their limb and motor function over time. Without proper guidance and feedback, patients may abandon their treatment due to the risk of secondary injury. Automated assessment of movement quality, employing skeletal movement data obtained from depth imaging devices, has the potential to enhance home-based rehabilitation by offering crucial quantitative feedback. Inspired by the success of transformer-based models with self-attention mechanisms in natural language processing and computer vision domains, we propose a transformer-based network for assessing movement quality, aiming to address the challenges associated with movement assessment in rehabilitation. Furthermore, our proposed model has been augmented with self-attention and channel attention mechanisms, enabling the transformer network to selectively attend to specific anatomical regions that play a crucial role in motion execution. Compared with traditional convolutional neural networks and long short-term memory, the proposed models obtain the best accuracy on two open-access datasets. And our contactless method also provides a new potential tool for automated movement assessment and telemedicine.