The precise recognition of gymnastics movements is of great significance for optimizing athlete training, ensuring fair competition evaluation, and promoting the scientific development of gymnastics. However, traditional action recognition methods are limited by the singularity of algorithm structure, making it difficult to effectively capture the high complexity and spatiotemporal dynamic characteristics of gymnastics movements, resulting in insufficient recognition accuracy and limited generalization ability. A gymnastics action recognition model combining Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) is proposed to address this issue. Batch normalization technology is introduced to optimize network training stability, and feature fusion technology is used to integrate multi-level spatiotemporal features to enhance the model’s expressive ability and robustness. The experimental results show that the optimized model performs well on the Sports-1 M and KTH datasets, with recognition accuracies of 96.4% and 94.1%, respectively, and average recognition times of only 3.7 ms and 3.4 ms. In addition, the model outperforms the comparison method in terms of recall rate (0.95 and 0.98) and F1 value (0.93 and 0.94). The case analysis further validates the recognition ability of the model for six types of gymnastics movements, including walking, running, jumping, balancing, spinning, and tumbling, with accuracy rates exceeding 85%. The model maintains stable performance under different movement speeds (slow movement mAP 91.4, fast movement mAP 88.7) and lighting environments. The research not only provides efficient and reliable technical solutions for gymnastics movement recognition, but also lays a theoretical foundation for real-time analysis in complex sports scenarios, which has important academic value and application potential.
Ran et al. (Wed,) studied this question.