CLASSIFICATION OF RHYTHMIC GYMNASTICS SPORT ELEMENTS BY VIDEO

Abstract

The work devoted to human posture recognition during rapid movements and complex non-standard poses due to the large number of limbs involved in the movement. Rhythmic gymnastics was chosen as the subject area and, accordingly, the specifics of the judge's assessment of the athlete's performance. Many synchronized videos with fast movements and sequences of complex poses from different angles allows us to form a data set necessary for further research and implementation of the results obtained both in socially important industries and in the market of commercial services using artificial intelligence technologies. A computer system has been developed that can be used to increase the objectivity of sports judging at rhythmic gymnastics competitions, as well as to become an alternative to the traditional judging system in the case of competitions held in a remote format. By scaling up the task, the system can also be used to diagnose problems with the human nervous system and musculoskeletal system. As a result of the research, a dataset depicting the performance of sports elements was collected and structured. The peculiarities of the mediapipe and ViTPose models were identified and the best solution for preprocessing the prepared set was chosen. The main result of this work is a built and trained model for classifying sports elements, which classifies 7 elements with an accuracy of 0.9048. The accuracy indicates that the model performs at a high level, correctly classifying sports elements in most cases. This level of accuracy indicates that the model has been effectively trained to classify these specific elements. In the future, to be able to fully evaluate the performances of female rhythmic gymnasts, it is necessary to add tracking of the object with which the athlete performs, to create a method for tracking interaction with it.