Real-Time Yoga Pose Detection and Classification Using CNNs for Enhanced Fitness and Rehabilitation
Banrilin Syiemlieh, Lapynhunshisha Jyrwa, Chemerita Ch Marak, Sarat Kumar Chettri · 2025
The growing demand for yoga pose detection has revolutionized health monitoring, fitness training, and rehabilitation practices. In response, we propose a yoga pose detection and classification system powered by deep learning, specifically utilizing Convolutional Neural Networks (CNNs), to achieve accurate, reliable, and real-time posture tracking. CNNs are particularly effective due to their ability to extract critical features from images, enabling recognition of various angles associated with specific yoga poses. Our system is designed in two phases: first, by training on a dataset comprising labelled images of common yoga poses, and second, by integrating real-time video analysis to provide immediate feedback on pose accuracy. To enhance model performance, we augmented the dataset to account for diverse body types, backgrounds, and pose variations. Additionally, we incorporated pre-trained models such as VGG16 for transfer learning, leveraging prior knowledge to improve generalization, reduce training time, and achieve higher accuracy. The system focuses on key yoga asanas, including Chair, Goddess, Tree, Downward-Facing Dog, and Warrior One poses. Its real-time functionality dynamically analyzes video frames from the user’s camera feed, extracting essential body landmarks and joint angles. These are then compared to an expert pose, enabling the system to evaluate user alignment and deliver corrective feedback for better pose execution. By combining robust dataset-based training with real-time video analysis, this approach demonstrates the potential of CNNs in supporting personal fitness. It provides accessible, automated yoga pose recognition and correction, making precise pose adjustments available for at-home users.