DESIGN OF SPORTS ACTION RECOGNITION AND EVALUATION BASED ON IMPROVED DTW ALGORITHM

With the popularization of sports and the development of computer technology, the demand for sports action recognition and evaluation is increasing day by day. Although existing methods have achieved certain results, there are still shortcomings in recognition accuracy, real-time performance, and stability. To improve the effectiveness of sports action recognition and evaluation, this study proposes an action recognition and evaluation method based on an improved dynamic time warping algorithm. It utilizes an improved 3D convolutional network (C3D-Resnet) to extract sports action features, and combines feature fusion and dimensionality reduction methods to improve the dynamic time warping algorithm. The test results on the CASIA TaiChi Dataset showed that the accuracy of C3D, Resnet, and C3D-Resnet were 84.5%, 86.4%, and 94.7%, respectively. After feature data dimensionality reduction, the data dimension decreased from [289, 678] to within the range of [17, 109], and the average action recognition rate increased from 88.3% to 91.2%. The F1 value of the improved dynamic time regularization algorithm was about 97.1%, and the difference between the sports action evaluation results and the scores of professional coaches was less than 1 point. This study has achieved accurate recognition and evaluation of sports movements, which has important application value and practical significance in improving the effectiveness of sports training and teaching.