Automated Yoga Pose Classification Using Deep Learning on Image-Based Datasets

The With the increasing popularity of yoga as a form of physical and mental fitness, the correct recognition of yoga poses is crucial for automated feedback systems, virtual training and health monitoring, and other applications. The main contribution of the present work is to introduce a yoga pose recognition pipeline based on deep learning methods for image-based datasets. The model makes use of MobileNetV2 for posture classification and YOLOv5 for real-time pose detection, which allows the system to perform efficient feature extraction and localize the body joints spatially. The models were trained and tested using a large dataset of multiple poses of different types and backgrounds in yoga. As experimental results, MobileNetV2 also had a good validation accuracy of 72.6%, also being able to effectively classify postures with a high frequency, while YOLOv5 also obtained a fairly good overall accuracy of 70%, being able to robustly detect and classify it at the same time. It can be concluded that in a single frame, where multi-pose detection is required, YOLOv5 is more suited, while for isolated posture recognition MobileNetV2 performs better for isolated posture recognition with MobileNetV2. This framework emphasizes the promise of deep learning for developing yoga intelligent training systems that can give automatic, real-time feedback to attendees, for applications focused on health and rehabilitation purposes.