Classification of Kyphosis Rehabilitation Exercises Using MediaPipe and Hybrid CNN-RNN Architecture

This paper introduced a deep learning-based approach for the automated classification of kyphosis rehabilitation exercises. This paper introduces models that are capable of analyzing recorded human motions. The proposed models are capable of classifying the correct and incorrect performance of the considered exercises. A custom dataset comprising over 500 video recordings of five k yphosis e xercises w as c reated under the supervision of a professional physical therapist. Utilizing MediaPipe, 33 different joint positions in 3D space are extracted from each video resulting in 99 features per frame which are then fed into deep learning models for classification. The effective performance of each Recurrent Neural Network (RNN), in particular Gated Recurrent Units (GRU) and Long Short-Term Memory (LSTM), is investigated. Moreover, their hybrid architectures with Convolutional Neural Networks (CNNs), namely CNN-LSTM and CNN-GRU, were studied for this classification task. Upon comparison, this paper proves that CNN-GRU outperforms other classifiers i n t erms o f a ccuracy, p recision, r ecall a nd F1 score. The use of CNN-GRU has proved to enhance the classification accuracy to reach 99.01% with an achieved precision of 99.09% and F1 score of 99.01%.