3D Yoga Pose Detection and Classification using Machine Learning Libraries

3D Yoga Pose Detection and Classification using Machine Learning Libraries

1Sarthak Mahendroo, 2Shruti Gupta, 3Harsh Rohilla, 4Amit Yadav, 5Ms. Gurpreet

1,2,3,4 B.Tech. Students, CSE Department, HMRITM, New Delhi, India

 5 Assistant Professor, CSE Department, HMRITM, New Delhi, India

B. Tech 4th Year, Dept. of Computer Science and Engineering,

HMR Institute of Technology and Management, Hamidpur, Delhi-110036

ABSTRACT

India has a lengthy history of being connected to yoga, an age-old art form. It helps a person's body to be healthy and at the same time provides peace of mind. With the emergence of Covid-19, doing yoga in a class full of people has become difficult and can cause serious injuries if done without guidance on the other hand, here we identify the different yoga poses for a user to perform by developing a web application. The application uses open-source data containing images of 5 different yoga poses performed by different volunteers. OpenCV handles all the images for this application. The system has two phases, the first extracts points of data from the image dataset using the MediaPipe pose estimation library, the second phase pre-processes the acquired data with points and performs classification-based training and tests the data using machine learning algorithm. The machine learning algorithms used are Logistic Regression, Random Forest Classifier, Gradient Boosting Classifier, K-Nearest Neighbour Classifier, and Ridge Classifier. The system achieves an accuracy score of about 98%. The application is designed to process images with a threshold of still images and live video, so solutions below a certain score are unacceptable.

Human estimation is a difficult problem to solve in the domain of computer vision like locating human joints in an image or video and creating a skeletal representation. Automatic recognition of human pose in images relies on many aspects, such as image scaling and resolution, lighting changes, background noise, clothing changes, environment, and human- environment interactions, hence making it a difficult task. An application of human pose estimation that has intrigued many researchers in this field is exercise and fitness. This is an ancient practice that began in India, but is now world renowned for its many mental, physical and spiritual benefits. However, the problem with yoga, as with any exercise, is that improper posture during a Yoga session can be counterproductive and potentially harmful. And that's why you need an instructor to correct your posture. Not all users have access to instructors or instructing resources. However, one may use artificial intelligence-based applications to identify yoga poses and provide personalized feedback to improve personal modifications. In recent years, human pose estimation has benefited from deep learning, which has significantly improved performance. Machine learning approaches offer a simpler way to map structures than dealing with dependencies between structures. Using machine learning and deep learning, we identified 5 yoga poses namely tree pose, plank pose, downward dog pose, warrior-2 pose and goddess pose.

Keywords: OpenCV, MediaPipe, Pose detection, Gradient boosting.