YOGA POSE DETECTION USING DEEP LEARNING

Abstract

Yoga, a centuries-old practice promoting physical and mental well-being, has gained global popularity, creating a need for technological solutions to assist practitioners in performing poses correctly. This research presents a deep learning-based yoga pose detection system that classifies images into various yoga postures to support real-time feedback and posture correction. We conducted a comprehensive comparison of five deep learning models: MobileNetV2, VGG16, ResNet50, EfficientNetB0 (all implemented through transfer learning), and a custom Convolutional Neural Network (CNN). Our experimental results reveal that MobileNetV2 achieved the highest classification accuracy of 82.42% while maintaining the lowest training time of approximately 451 seconds, indicating its suitability for deployment in resource-constrained environments such as mobile applications. In contrast, VGG16, despite achieving a moderate accuracy of 71.48%, required significantly longer training times exceeding 10,000 seconds, limiting its practical usability. ResNet50, Custom CNN, and EfficientNetB0 yielded notably lower accuracies of 12.5%, 8.20%, and 6.25%, respectively, suggesting these architectures may require further optimization or more extensive datasets for this application domain. The study underscores the critical balance between model complexity, computational efficiency, and accuracy for practical yoga pose detection systems. Our findings demonstrate the promise of lightweight transfer learning architectures in achieving efficient and effective yoga pose recognition, paving the way for smart fitness applications capable of delivering personalized guidance and injury prevention. Future work will explore data augmentation strategies, real-time video-based detection, and user-specific customization to further enhance performance and usability in real-world settings.