Action Recognition and Sports Evaluation of Running Pose Based on Pose Estimation

At present, deep convolutional neural networks (CNNs) have made impressive progress on human pose estimation. Sports pose is one of the important bases for evaluating the athletes' skill level. Until now, the determination and analysis of athletes' pose is still very time-consuming manual work. Therefore, how to adapt and retrain the state-of-the-art human pose detection system and apply it to specific sports competitions is of great significance. Videos recorded on a regular basis can quickly provide a large amount of pose data, and the video recording work is a preliminary preparation for the development of mining algorithms. In this study, we use running as an example of a sport. Using the High-Resolution Net (HRNet) neural network algorithm, we obtain 2D pose data from the video recordings of individual movements to realize the recognition and evaluation of athletes' movement pose during running. This breaks the research boundaries of traditional laboratories and realizes the analysis of movements during actual running competitions. Our experimental results showed the effectiveness of this method, which has the potential to be applied to other cycle-based or phase-based types of sport. The exploration of this study has achieved the identification of running motions and the analysis of relevant motion variables, such as joint angle, stride frequency, stride length, etc. This provides a more scientific technical basis for us to adopt this method in practical sports video analysis in the future.