Dance Therapy Motion Recognition and Quality Evaluation Based on Hybrid Dilated Convolution and Spatial-Temporal Aggregation Network

Dance therapy is a commonly used rehabilitation training, and the quality of patients' movements affects the training effect. With the rapid development of artificial intelligence (AI), various motion recognition and quality analysis methods have emerged. This study proposes an innovative design: a motion recognition method named Spatial-Temporal and Channel-wise Aggregation with Multi-Scale Network (STC-MSN) that introduces hybrid dilated convolution (HDC), and a motion quality evaluation algorithm based on improved Fast Dynamic Time Warping (FastDTW). The STC-MSN motion recognition model is designed. This model integrates four components: temporal aggregation, channel topology modeling, spatio-temporal channel fusion, and multiscale temporal convolution, which can capture the spatial-temporal features of skeleton movements. Experimental results show that compared with existing methods, the proposed method has remarkable advantages. The top-1 accuracy of the STC-MSN model on the National Taiwan University RGB-Depth (NTU RGB+D), NTU RGB+D 120, and Kinetics-Skeleton datasets is 94.77 %, 86.68 %, and 35.21 %, respectively. The Spearman correlation coefficient ρ of the improved FastDTW on the University of Idaho Physical Rehabilitation Movement Dataset (UI-PRMD) and the Kinect-based IMU-enhanced MOnitoring of REhabilitation (KIMORE) dataset is 0.821 and 0.794, which is higher than that of the baseline model. Meanwhile, its mean squared error and mean absolute error are the smallest. These results indicate that the algorithm has high motion recognition accuracy, good stability, and high-quality evaluation accuracy. When applied to dance therapy, it can perform accurate motion monitoring and timely correction for patients, customize training plans, and improve training effects and patients' movement completion.