Analyzing Yoga Pose Recognition: A Comparison of MediaPipe and YOLO Keypoint Detection with Ensemble Techniques

The practice of yoga encompasses mental, physical, and spiritual dimensions, aiming for holistic wellbeing. Accurate alignment in yoga enhances the effectiveness of each pose by targeting specific muscle groups, reducing strain on muscles and joints, and improving stability and balance. This research employs advanced computer vision techniques, YOLO (You Only Look Once) and MediaPipe to identify critical keypoints from the skeletal structures of yoga practitioners, thereby providing a detailed representation of body alignment for posture recognition. Augmented using the SMOTE technique, the skeletal data serves as input for various Machine Learning and ensemble models during the training process. The study utilizes a 2D image dataset comprising 20 well-known yoga poses. Among the models tested, the LightGBM ensemble classifier using MediaPipe keypoints achieved the highest accuracy at 96.52%. Further analysis included the evaluation of the model through a confusion matrix, learning curve, and pose-wise accuracy, even for similar-looking exercises. These findings highlight the potential of integrating computer vision and machine learning to enhance yoga practice through precise posture recognition and alignment analysis.