Yoga, an ancient discipline emphasizing balance, flexibility, and mindfulness, relies heavily on maintaining correct pos- tures for achieving physiological and psychological bene- fits. However, for most practitioners, accurately performing these poses without expert supervision remains challenging. With the advancement of computer vision and deep learn- ing, automatic yoga pose recognition has become an emerg- ing area of research aimed at assisting practitioners through intelligent feedback systems. Traditional approaches pri- marily depend on convolutional neural networks (CNNs) trained directly on image pixels, which, while effective, require large datasets and are computationally expensive. Moreover, such models are often sensitive to background variations, lighting conditions, and camera angles. To overcome these challenges, this work presents a lightweight and interpretable method for yoga pose classifi- cation using MediaPipe Pose for keypoint extraction and a deep neural network (DNN) trained on derived joint-angle features. By representing each posture through geometrical relationships among body landmarks instead of raw images, the proposed model achieves higher generalization across different environments while maintaining real-time perfor- mance. This approach not only reduces computational com- plexity but also provides a foundation for developing inter- active yoga assistants capable of guiding users toward cor- rect pose alignment. The novelty of this work lies in the use of geometric, angle-based pose representation derived from human body landmarks, instead of raw pixel data used in CNN-based methods. This approach makes the model lightweight, interpretable, and background-invariant while maintaining a high accuracy of 94%. Furthermore, the pro- posed architecture achieves real-time performance on CPU hardware, making it suitable for low-resource yoga assis- tance and corrective feedback systems.
You can download the full project report here: