Riska Dhenabayu, Nanang Hoesen Hidroes Abbrori, Hujjatullah Fazlurrahman, Achmad Fitro
This research explores the application of hybrid transformer networks for video-based action recognition in sports specifically for badminton. This research aims to capture and classify badminton player movements with high accuracy. The novelty of this study lies in its hybrid transformer approach, a sophisticated blend of transformer architectures and self-supervised pretraining tasks, designed to process sequential image data, leveraging the spatial-temporal capabilities of transformer models to enhance feature extraction and reduce computational complexity. Initial experiments were conducted using a curated VideoBadminton dataset, achieving 86.2% CS accuracy and 92.9% CV accuracy, a promising increase in precision compared to traditional convolutional neural networks. This breakthrough indicates that hybrid transformer networks can significantly improve the accuracy in video-based badminton action recognition. The findings suggest a scalable path forward for integrating advanced AI methodologies into sports analytics for AI driven decision support system, offering substantial benefits for performance assessment and athlete development. © 2024 IEEE.
Universitas Negeri Surabaya, Faculty of Economics and Business, Surabaya, Indonesia
