Physical Rehabilitation Exercises Classification Using Deep Learning Models

Abstract

Physical rehabilitation plays pivotal role for restoring functionality and enhancing well-being in individuals with injuries, surgeries, or illnesses. This study introduces a framework to monitor patient progress during rehabilitation and identify the parts of skeleton for each exercise involved in each exercise. First experiment using UI- PRMD dataset, accuracies were 98.50%, 87.8%, 92.4% for LSTM, CNN-LSTM and GRU respectively. The second experiments with a collected dataset showed accuracies of 98.11% for LSTM, 71.7% for CNN-LSTM, and 96.23% for GRU, with additional promising results from DenseNet models and 3D array representations.