<![CDATA[Practicing yoga independently without professional supervision can lead to incorrect postures, increasing the risk of injury and reducing exercise effectiveness. Although various studies have utilized pose estimation and machine learning techniques for yoga pose classification, most focus on a single feature representation. This study proposes a static image-based yoga pose classification system using MediaPipe to extract 33 body landmarks, which are transformed into geometric features consisting of landmark coordinates, joint angles, and inter-body point distances. The novelty of this study lies in the systematic evaluation of these features, both individually and in combination, within a Random Forest classification framework. The dataset consists of 1,531 images representing five yoga pose classes: Downward Dog, Goddess, Plank, Tree, and Warrior II. The Random Forest model was optimized using hyperparameter tuning and cross-validation. Experimental results show that combining landmark, angle, and distance features achieved the best performance, with an accuracy of 95.02% and an F1-score of 0.9501. The model also demonstrated stable performance during cross-validation, with accuracy ranging from 0.9436 to 0.9608. The results show that combining multiple geometric feature representations improves yoga pose classification performance while maintaining computational efficiency, supporting safer and more effective self-guided yoga practice]]>
Copyrights © 2026
