Neural network methods for automatic person pose estimation in rhythmic gymnastics exercises

Abstract

In rhythmic gymnastics, evaluating the performances of female athletes is a complex subjective task due to the need to consider both the technical characteristics of dynamic exercises and the aesthetic perfection of individual scenes and the composition as a whole. The article presents the results of a study aimed at developing a program for automatically detecting the postures of female athletes during rhythmic gymnastics exercises. The development of the model is preceded by a study of the current scoring rules, considering the complexity of the athlete's movements. For the study, a special dataset was created to determine the gymnasts' poses, which included scored elements and unscored positions. The dataset contains freeze frames from the video of the competition and photo reports. Two computer vision methods are used to identify the poses: MediaPipe Pose and ViTPose. MediaPipe Pose allows for real-time detection of 33 3D landmarks due to its high performance. ViTPose provides high accuracy using visual transformers. A comparative analysis of these two methods is carried out, highlighting the strengths and limitations of each of them. A multilayer perceptron model has been developed to classify sports elements based on recognized poses. The model was trained using back-propagation of error based on Adam's gradient descent method. After training the model on the author's dataset, high accuracy in the classification of gymnasts' poses is achieved. This study is an intellectual component of a future system for automatic real-time detection of an athlete's pose for a more reliable and efficient evaluation of rhythmic gymnastics performances. The proposed approach to combining computer vision and machine learning methods can be extended to improve sports analysis in other related disciplines.