Machine-Learning Technique For Camera-Based Monitoring And Evaluation Of Yoga Posture

Abstract

Human pose estimation is a long-standing issue in computer vision that has presented numerous difficulties in the past. Robotics, Video surveillance, biometrics, Augmented reality, assisted living, at-home health monitoring, and other industries benefit from analysing human actions. This project establishes the groundwork for developing a feedback-evaluation system by combining a machine learning-based model with deep learning methodologies to categorise yoga positions in both pre-recorded and real-time footage.
Our project is a browser based system divided into 3 parts - Creating the dataset, Training the model, Classification of pose for 3 Yoga asanas using the poseNet and neuralNetwork models which are present in the ml5.js library. PoseNet is a pre-trained model which detects 17 keypoints on the human body when input is provided through pre-recorded videos or in real-time using a webcam. These keypoints are stored in the form of x,y locations and used to train the overall system using the neuralNetwork model. The classification process implements both poseNet and neuralNetwork models simultaneously. The user's pose is compared to the expert's pose, and the angle difference between various body joints are determined.