Deteksi Ketepatan Pose Yoga menggunakan Metode Bidirectional Gated Recurrent Unit Berbasis Video

Abstract

Yoga is a physical and mental activity that is gaining popularity among people, especially the younger generation, for its benefits in improving physical health and emotional balance. However, practicing yoga on one's own without an instructor poses a risk of injury if the poses are incorrect. This problem is exacerbated by limited access to professional yoga classes due to time and cost constraints. This research aims to develop a yoga pose accuracy detection system based on deep learning technology using the Bidirectional Gated Recurrent Unit method. The system is designed to provide automatic feedback and pose correction to users who perform yoga exercises independently. The dataset used in this research consists of six classes of yoga poses, namely Downward Dog, Tree, Plank, Boat, Warrior II, and Shoulder Stand, which are obtained from Youtube videos and personal recordings. The video data was then extracted into frames using OpenCV and processed using the MediaPipe framework to generate body keypoints. The keypoints data is formed into a sequence and used as input for the BiGRU model to detect overall pose accuracy based on movement dynamics. The implementation results show that the system is able to detect poses with 93.3% accuracy and provide feedback in the form of voice and text alerts if the user's movements are not appropriate. Thus, this system can help users practice yoga more safely, effectively, and efficiently without the need for the presence of a professional instructor. This research is expected to be an alternative artificial intelligence-based solution in supporting a healthy lifestyle and reducing the risk of injury due to posture errors in independent yoga practice.