Human Pose Estimation for Boxing Based on Model Transfer Learning

The boxing as a competitive sport is now walking into Chinese campus and daily fitness exercises. The rapid growth of boxing sports has resulted in a coaching shortage. The pose of boxing can be estimated by using human pose estimation technology in artificial intelligence, to teach interns and relieve the coach shortage. However, the large number of training datasets and the lack of consistency of camera views in training limit the technology's practical application. Therefore, the model transfer methodology was implemented to address these limitations, three basic models of OpenPose (OP), High Resolution (HR), and stacked Hourglass (HG) networks were employed for model transfer learning and analyzing of boxing application. The results show the model transfer can efficiently improve the average accuracy of pose key points from 1% to 19%, model transfer can practically help address the issue of dependence on dataset collection. The HR network achieves the best performance among these three networks, but the HG network still gets better estimation on some key points.