Enhancing AI-Driven Sports Dance Therapy via Hybrid Dilated Convolution: A Motion Recognition and Evaluation Optimization Framework

This study addressed the need for motion process monitoring and real-time feedback in AI-assisted sports dance rehabilitation training. It proposed a multi-scale temporal evaluation framework incorporating Hybrid Dilated Convolution (HDC). Unlike traditional single-scale temporal models, this study addressed the difficulty of jointly modeling local rhythm variations and long-term motion structure. It introduced convolutional branches with different dilation rates and stage-aware cross-scale fusion and incorporated temporal attention and a joint optimization strategy for action recognition and quality scoring. This design enabled collaborative modeling of short-term rhythm deviations, motion transition features, and long-sequence structural dependencies. The core innovation lay in introducing action phase information into multi-scale feature extraction and the scoring process. It combined rhythm consistency constraints with stability constraints. This formed a process-level evaluation mechanism for coordination, continuity, and motor control quality. The experiment was conducted using the publicly available Artificial Intelligence and Society Team (AIST)++ dance dataset for methodological validation. The dataset contained approximately 5.2 hours of paired dance and music data. It included a total of 1,408 motion sequences and covered 10 dance styles. It also provided multi-view videos, human keypoint sequences, and musical beat information. The dataset supported action recognition, rhythm alignment, and temporal quality analysis. On the rhythm-enhanced subset, the proposed model achieved an action recognition accuracy of 0.932. The Matthews correlation coefficient reached 0.906, and the Kappa coefficient reached 0.889. For action quality scoring, the mean absolute error and root mean square error were 0.186 and 0.253, respectively. On the motion-transition-dominant subset and long-sequence subset, the model also maintained stable performance in both recognition and scoring tasks. Efficiency evaluation showed that the model achieved an inference latency of 13.284 ms and a throughput of 75.936 FPS. The peak memory usage was 1,738.294 MB. The parameter size was 28.637 M, and the computational cost was 11.347 GFLOPs. The results indicated that the proposed framework had strong technical feasibility in multi-scale temporal motion recognition, process-level quality evaluation, and real-time feedback. It can provide a reference for motion monitoring and assisted feedback in sports dance rehabilitation training. This method still requires further validation in clinical rehabilitation settings. Such validation should involve real patient data, therapist evaluations, and standardized clinical scales.