Athlete movement control training based on dynamic time warping algorithm

• Published in: Systems and soft computing Vol. 7; p. 200336
• Main Authors: Dai, Luo, Zhao, Yueqiang
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2025; Elsevier
• Subjects: Action assessment; Action recognition; Dynamic time regulation; Long short-term memory network; Spatial-temporal graph convolutional network; Action assessment; Action recognition; Dynamic time regulation; Long short-term memory network; Spatial-temporal graph convolutional network
• ISSN: 2772-9419; 2772-9419
• DOI: 10.1016/j.sasc.2025.200336

The training process of athletes' movement control usually relies on the subjective judgment of the coach, making it hard to quickly and accurately analyze the standard degree of movements. Given this, this study optimizes the training process of athletes' motion control based on spatiotemporal graph convolutional networks and dynamic time-warping algorithms. Firstly, an athlete action recognition model based on a spatiotemporal graph convolutional network is established, and a recognition termination strategy based on LSTM is designed by using the non-maximum suppression method to extract redundant frames. Subsequently, a quality assessment model for action completion is built based on an improved dynamic time-warping algorithm. Finally, the performance of the proposed model is studied in the PennAction dataset, KTH dataset, HiEve dataset, and the collected wrist movement data of the subjects. The results indicated that the proposed action recognition model had a recognition accuracy of 93.15 % in the PennAction dataset, a gigabit floating-point operation of 25.7, and a loss value below 0.02. The recognition accuracies on the KTH dataset and HiEve dataset were 95.07 % and 91.54 %, respectively. Improving the dynamic time-warping algorithm avoided horizontal or vertical distortion during the action-matching process. The scoring results of the proposed action completion quality evaluation model basically followed a normal distribution, with an accuracy of 92.16 % and a time consumption of 1.73 ms, which was 0.27 ms shorter than traditional algorithms. The experiment has demonstrated the motion recognition and matching performance of the proposed model. The research results help reduce subjective judgment errors in athletes' motion control training and improve training effectiveness and competition results.