Yoga Pose Classification and Correction Using Deep Learning

Abstract

"The author proposes a solution to address the challenge of maintaining correct yoga postures, which are essential for the physical and mental benefits of this Indian cultural exercise and meditation method. Incorrect postures can lead to serious harm and long-term joint pains, making it crucial to ensure that individuals perform yoga poses correctly. The solution presented is the development of a yoga assistant using pose estimation technology to classify and correct user postures.

To utilize this system, users can either upload a video of themselves performing yoga poses or engage in real-time posture correction. The system extracts frames from the video and identifies key points on the user's body, subsequently classifying the performed posture and calculating any errors. This enables users to receive immediate feedback on the accuracy of their postures. The system utilizes the LCRN model for posture classification and the MediaPipe deep-learning library for correction by calculating the angles between key body points.

The testing and evaluation of the system involved domain and technical experts, as well as general users. Testing encompassed feedback from users across different domains, ensuring the system's usability and effectiveness. Technical evaluation included metrics like confusion matrices, model loss graphs, and accuracy assessments, all of which yielded satisfactory results. Additionally, the system underwent black-box testing, further confirming its reliability.

In summary, the proposed solution involves the development of a yoga assistant that uses pose estimation to classify and correct yoga postures, providing users with a valuable tool to enhance their yoga practice and ensure its effectiveness. This technology has been rigorously tested and evaluated to ensure its accuracy and usability. "