Poses Classification in a Taekwondo Lesson Using Skeleton Data Extracted from Videos with Shallow and Deep Learning Architectures

Abstract

AbstractSports is an important activity to help maintain and improve human health, help fight diseases, create flexibility for the body, and contribute to training the spirit of competition, spirit teammates, and increase soft human skills. The current sports practice has great support from technical technology, contributing to sports to achieve many high achievements. Currently, many technological techniques are applied to support sports for practicing and monitoring matches. For example, Video Assistant Referee re-examines videos to support the referee’s accurate decisions on a specific situation in football matches and applications in poses recognition of aerobics and martial arts sports. This study uses Fast Forward Moving Picture Experts Group (FFMPEG) technique to extract images from the video of Taekwondo and generate skeleton data from extracted frames with MoveNet. Then, we perform the poses classification tasks for TAEGEUK IN JANG lesson with deep learning architectures such as shallow convolutional neural networks, VGGNet, Inception, and Long Short-Term Memory networks. These architectures are modified to receive 1-Dimensional data as input, including key points of the skeleton. Poses recognition tasks in sports lessons use skeleton data to eliminate noise in the images, such as background behind practitioners and unrelated objects, and only focus on the movement/direction of the poses. Our proposed method has achieved good accuracy in the data, including 35 videos (more than 25,000 frames) distinguishing 20 poses in a basic Taekwondo lesson.KeywordsTaekwondo lessonPoses classificationAction recognitionSkeleton