Сучасні технології для визначення ланок тіла борців

Abstract

Purpose: to determine the effectiveness of using modern computer vision technologies to determine the body parts of wrestlers in competitive activities. Material and methods. The following methods were used in the research: analysis of scientific and methodical literature and Internet sources, analysis of video and images with the help of computer vision (Computer Vision, Machine Learning), methods of mathematical statistics. The final bout in the 65 kg weight category at the 2021 Tokyo Olympics in freestyle wrestling was chosen for analysis. The total number of video fragments was 1763 images. The Yolo machine learning model was used to detect faces, and Google's MediaPipe model was used to detect body parts. For each athlete, 21 body parts were identified. Results: the accuracy of identifying athletes' body parts was obtained: all indicators were above 78 %. The results of the analysis showed that the lower limbs are better identified: hip, knee, ankle, etc. The identification accuracy is above 90%. At the same time, the identification accuracy of the upper limbs is lower. In our opinion, this is due to the fact that athletes spend a sufficient amount of time in the grip and preparing for an attack during the fight. During this time, certain parts of the upper limbs are closed for identification. Analysis of the results of the number of unidentified images showed a similar trend to the identification accuracy. Most of the unidentified body parts were associated with the upper limbs (more than 25 %). At the same time, 10 to 12 % of the body parts associated with the delicate limbs were not identified. Conclusions. The analysis of results of the use of modern technologies of computer vision for determination of links of a body of wrestlers in competitive activity is carried out. The algorithm for studying the body parts of athletes had two stages: the first stage is to determine the area of the image fragment (video) where the athletes are located and the second is to determine the body parts of each wrestler. At the first stage of the study, it was possible to identify both athletes in 79,5 % of the images. The analysis of the accuracy of the athletes' body parts identification showed that the body parts of the lower limbs are better identified than those of te right limbs. This is due to the peculiarity of competitive activity. The visual analysis of the images showed that a certain part of the identified links do not coincide with the actual location of the athletes' body links.