Taekwondo action recognition model based on deep learning and key point matching technology

In order to solve the problems in Taekwondo action recognition and key point matching, this paper proposes a hybrid matching model that combines YOLOv8 and an improved dynamic time warping (DDTW) algorithm. The model uses YOLOv8 to detect key points in each frame of the video. When the feature points are lost due to occlusion, the model tries to use the information of adjacent frames to compensate. At the same time, in order to reduce jitter, the key point coordinate sequence is smoothed by a low-pass filter. The key frames are matched by the DDTW algorithm, and the action score is calculated according to the best path found by the algorithm. The scoring basis includes the cumulative cosine similarity, action fluency and accuracy. The experimental results show that by combining the images of the three perspectives, the model has achieved significant improvement in recognition accuracy, especially in the key frame combination extracted at the quartile position, the highest recognition accuracy of the front view and the right view reached 92.1% and 94.3% respectively. The method shows strong robustness in noise and speed change processing, and is suitable for motion analysis, rehabilitation training and sports science.