Deep Learning-based Accurate Recognition and Analysis of Saber Movements

This study aims to explore and implement a deep learning based saber movement recognition and analysis method, to enhance the understanding and mastery of saber sports technical skills. It gets significantly uses in saber sports, where accurate movement identification and extraction are fundamental for technical training and strategy improvement of athletes, and ensure unbiased decisions of referee in computation. Traditional movement detection methods are heavily rely on visual methods, which are susceptible to subjectivity and variable performance. We make a modified version of the YOLOv8x-Pose model, and improves the logistic regression model with an OVR strategy, to boost the performance of recognizing saber actions in a fast moving and dynamic environment. We introduce a self-built dataset of 2,040 images of basic saber actions, to increase training efficiency and improve classification performance of the model. By data preprocessing, feature extraction, and model training. The experimental results shows that the classification latency with the optimal model in real-time video streams is less than 0.5 seconds and the accuracy is 93.2% in total. Finally, the study elaborates on future such expansion of data set, improvement of robustness of the models, and reusable application of the deep learning modeling in analysis of saber actions. The article puts forward a new technical means for doing scientific training and competition strategy in saber sports with great practical significance and application value.