This study aims to address the insufficient utilization of second-order information in current skeleton-based action recognition techniques for sports dance, thereby meeting the high-quality rehabilitation requirements of modern sports dance training. The proposed approach employs the AdaScan action extraction method combined with the OpenPose tool for human key point detection, effectively capturing critical features within dance movements. Additionally, a keyframe detection network based on AdaScan is used to accurately extract key video frames, while long-term temporal modeling is achieved through a Key Frame Segment Network (KFSN). Furthermore, a Convolutional Neural Network (CNN)-based action recognition system is developed to enable real-time monitoring of participants’ performance, allowing for the timely identification and correction of incorrect postures and providing a scientific foundation for physicians to formulate personalized rehabilitation plans. Extensive experiments are conducted on two large-scale benchmark datasets: NTU RGB+D and Kinetics. The results demonstrate that the proposed AdaScan-CNN model achieves an accuracy of 98.5%, a recall of 98.2%, and an F1-score of 98.3% on the NTU RGB+D dataset. On the Kinetics dataset, it achieves 96.7%, 96.5%, and 96.6%, respectively. Compared with conventional CNN models, the proposed method significantly improves accuracy, recall, and F1-score. These results strongly validate the superior performance of the AdaScan-CNN model in sports dance action recognition, providing more precise motion assessment and feedback for dance training and rehabilitation applications.
Wan et al. (Tue,) studied this question.