A Deep Learning Approach for Standing Yoga Pose Recognition using YOLOv11-based Keypoint Detection

The current research focuses on increasing the need for automated systems that enable real-time recognition of standing yoga poses because of the growing interest in yoga and smart fitness devices. Traditional machine learning algorithms rely on manually extracted features, whereas current state-of-the-art deep learning methods tend to use two different models – one for keypoint detection and another for classification of yoga poses. In order to overcome these problems, we propose a YOLOv11-based framework to detect and classify standing yoga poses in real-time. We collected a custom dataset with 2,250 images of nine different standing yoga poses using several people and various camera angles. Our algorithm applies YOLOv11 backbone, FPN-PAN neck, and specific pose head in order to perform effective keypoint detection and pose classification. We have experimentally shown that our method is better than YOLOv5 and YOLOv8. For a split of 70:20 train/validation, the achievements are 98.6% accuracy, 99.0% precision, 98.2% recall, 98.0% F1 score, and 98.6% mAP, whereas for a split of 80:10, they are 98.4% accuracy and 98.0% mAP.