Sports action estimation and recognition based on dynamic spatiotemporal graph convolution

To achieve accurate estimation and recognition of sports movements, an action pose estimation model was first constructed based on an improved graph convolutional network. Then, a motion recognition model was designed based on a dynamic spatiotemporal graph convolutional network. The experimental results showed that the motion pose estimation model designed in this study could detect a joint percentage of 0.986, and the success rate of joint point detection was 0.975. Compared with the real labels, the estimated joint points were the most similar, with a maximum average pose point position error of only 0.25. The action recognition model based on a dynamic spatiotemporal graph convolutional network achieved the best recognition accuracy and efficiency, with an average accuracy of 0.959, a frame rate of 192.988 FPS, and a computational load of only 98.754 FLOPs. Its Top-1 and Top-5 accuracies were as high as 0.96 and 0.97, respectively, indicating a strong ability to identify behavior from different perspectives and individuals. This study improves sports performance and advances sports research by accurately recognizing and estimating the movements of athletes.