A Graph Convolutional Siamese Network for the Assessment and Recognition of Physical Rehabilitation Exercises

Recently, due to the attention of physical rehabilitation improves markedly, several researchers attempt to implement automatic rehabilitation exercise analysis. However, most of the existing methods only focus on the assessment of a single action class, which limits the application scenario of multi-type action assessment. To advance the prior work, we present a novel graph convolutional siamese network to combine action classification and action assessment task. Specifically, a test action and a standard action form a pair as input to our model, which assesses the correctness of the test action compared with the standard action. Meanwhile, our model adopts a graph convolutional network to extract a feature from the input 3D skeleton data and recognize the action. Finally, we evaluate our model on UI-PRMD and IntelliRehabDS two popular datasets. Experiments demonstrate that the proposed model reaches state-of-the-art performance on action classification and outperforms the Dynamic Time Warping algorithm and hidden Markov model method by a large margin in terms of assessment accuracy.