YogAngleNet: Body angle-based yogic posture classification approach using MediaPipe from single-person images

The profound psychological and physical benefits of yoga have attracted global interest. This study explores yoga pose classification using angle values derived from single-person posture images. The dataset is carefully curated and consists exclusively of single-person yoga postures. The proposed method applies feature selection to improve model performance, achieving high accuracy with the random forest (RF) classifier. Using the top four selected features, our approach attains a maximum testing accuracy of 92.63%. A key contribution of this study is its strong focus on model interpretability. To explain model predictions and clarify decision-making, we employ Local Interpretable Model-Agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP). The results demonstrate the effectiveness of our approach, with potential applications in yoga instruction and analysis. This study advances yoga pose classification while highlighting the synergy among artificial intelligence, feature selection, and model interpretability, thereby extending the relevance of this ancient practice.