Enhancing Physical Therapy Through Transformer-Based Models: A Study on Exercise Classification

Abstract

Physical therapy is crucial for recovering and enhancing the physical functions of patients. It also improves the recovery process after injuries, surgeries, or diseases. Home-based rehabilitation has become essential in healthcare, particularly due to the limitations of traditional rehabilitation methods. This study proposes a system for monitoring patients, tracking their progress throughout rehabilitation, and identifying the skeletal points for each exercise. Also, identifying exercises based on deep learning single models, transformer models, and fused models using datasets collected by an expert physiotherapist through RGB cameras. LSTM, CNN-GRU, CNN-LSTM, GRU, and BiLSTM were applied with accuracies of 93.33%, 91.96%, 92.86%, 93.33%, 91.11% respectively. DenseNet201 achieved a higher of 96.42% and ViT-CNN with 91.71%. Furthermore, the human activity recognition transformer (HART) achieved an accuracy score of 91.96%