Research on Lower Limb Movement Rehabilitation Assessment Based on Graph Convolutional Network

In recent years, self-rehabilitation and assessment have become the primary choices in the mid-to-late stages of limb motion rehabilitation. However, self-assessment has a certain degree of subjectivity, leading to inaccurate quantitative evaluations that may not effectively guide the patient. Therefore, this paper constructs a more accurate quantitative motion rehabilitation assessment network based on human posture evaluation technology and graph convolutional networks, enabling the evaluation of lower limb rehabilitation movements. Firstly, this paper introduces a pLSTM self-supervised module, which is used to compute the joint loss function under a self-supervised mechanism, enhancing the limb motion assessment model's ability to learn richer feature representations. Secondly, an attention-guided drop mechanism is proposed for use in the temporal convolution, effectively achieving decoupling between channels and alleviating the issue of network overfitting. Finally, a new node partitioning strategy is proposed to construct the graph structure matrix, which better expresses the relationships between joints. The self-attention mechanism in the model is used to extract features from the joint graph, achieving a more accurate assessment that can better guide users in improving their training outcomes during evaluation. Experimental results show that our proposed LP-STGCN network model achieved significant improvements on the KIMORE dataset, with MAD improving by 0.192, RMSE by 0.2282, and MAPE by 0.4601. Compared with existing limb motion assessment methods, the model presented in this paper demonstrates higher recognition accuracy and serves as an important reference for further research in the increasingly important field of ambient assisted living.