Comprehensive Analysis of Pose Estimation and Machine Learning Classifiers for Precise Yoga Pose Detection and Classification
Meghana J.H., Chethan H.K., Kanhaiya Kumar, Suraj Prakash · Procedia Computer Science · 2025
Yoga contributes to mental and physical well-being by improving flexibility, strength, balance, and emotional stability when integrated into daily routines. This ancient practice can become more accessible and adaptable to a wider audience when combined with modern artificial intelligence (AI). This study introduces a comprehensive system for detecting and classifying yoga poses using computer vision and machine learning techniques. Central to this work is the application of posture estimation algorithms, such as MediaPipe, PoseNet, and OpenPose, to identify key points on the human body within a single image or video frame. These key points are analyzed in both two-dimensional (2D) and three-dimensional (3D) spaces to construct a skeletal representation of the body, enabling accurate classification of yoga poses. The study focuses on five distinct yoga poses: Downdog, Goddess, Plank, Tree, and Warrior II. To categorize these poses, machine learning classifiers including Support Vector Machines (SVM), Random Forest, K-Nearest Neighbors (KNN), and Naive Bayes utilize the key points extracted from the pose estimation models. This research is distinctive in its thorough evaluation of various conventional classifiers across multiple yoga positions. A comprehensive comparative analysis is essential for identifying the most effective classifiers for posture detection and classification. The dataset used in this study has been carefully curated to encompass a wide array of yoga poses and is divided into training and testing sets at various ratios (90:10, 80:20, 70:30, and 60:40) to ensure robust validation. Results indicate the system’s effectiveness, with SVM and KNN consistently achieving high values for Accuracy PE , Precision PE , Recall PE , and F1-S core PE across all yoga poses. Notably, Random Forest attains up to 100% Accuracy PE in detecting and classifying certain poses, demonstrating its robustness and reliability. This research highlights the potential of integrating pose estimation models with machine learning classifiers to create intelligent systems that can assist practitioners in yoga, marking an innovative step toward merging traditional practices with advanced technology.