PhysioFormer: A Spatio-Temporal Transformer for Physical Rehabilitation Assessment

Studies indicate that physical rehabilitation exercises recommended by healthcare professionals can enhance physical function, improve quality of life, and promote independence for physically disabled individuals. In response to the lack of immediate expert feedback on performed actions, developing an automated system for monitoring such actions and providing feedback is very much needed. In this work, we focus on skeleton-based exercise assessment, which uses skeleton data to evaluate human motion and provide a score on how well a patient performed a movement. There are several approaches to this issue, with Spatio Temporal Graph Convolutional Networks (GCN) being among the most recent. GCNs model skeleton data as graphs and utilize temporal and spatial convolutions to capture relationships between joints more effectively than previous methods. In this research, we propose a new Transformer based model, PhysioFormer. It is inspired by SkateFormer method for human action recognition, with enhanced structure to fit the task of physical rehabilitation assessment. The model leverages skeletal-temporal self-attention across different groups based on relations between joints. The evaluation is done on the KIMORE, UI-PRMD, and KERAAL datasets, benchmark datasets that provide skeleton data captured by Kinect motion system. Our model is surpassing state-of-the-art methods significantly.