Automatic Quality Grading of Rehabilitation Exercise with Skeleton-Based Deep Learning

Computerized examination of rehabilitation ex- ercises is crucial for tele-rehabilitation and physiotherapy. Current systems are based on costly depth sensors such as Kinect, which restricts availability. This work reviews Graph Convolutional Neural Network (GCNN) methods for exercise assessment based on human poses, drawing on both publicly available datasets (KIMORE, UI-PRMD, ExoRe- hab, NTU RGB+D) and a dedicated webcam-based dataset. GCNNs inherently capture the spatial joint relations of skeleton and temporal motion, supporting correct exercise classification. The hybrid dataset strategy provides real-time assessment with off-the-shelf webcams, a low-cost scalable option. Challenges, constraints, and directions for future tele-rehabilitation systems are explored.