Page 55 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Model Training and Fine-Tun- A convolutional neural network (CNN) model, MoveNet, was used
ing for processing videos generated while practising yoga poses. This 4.1-Healthcare
algorithm achieves a good level of accuracy and also provides
multi-person support. A custom dataset was prepared consist-
ing of 10 yoga poses: Bhujangasana, Garudasana, Halasana,
Natarajasana, Sukhasana, Makarasana, Savasana, Simhasana,
and Tadasana. In addition, the Yoga-82 dataset was studied [2].
Testbeds or Pilot Deploy- Active, Demo available on request
ments

Code repositories [1]

2 Use Case Description

2�1 Description

Yoga, recognised globally as an art and science for healthy living, has seen a significant surge
in popularity, underscored by the United Nations' declaration of June 21st as International
Day of Yoga. While its benefits are widely acknowledged, effective practice hinges on expert
guidance and consistent adherence. However, numerous challenges hinder widespread access,
including the difficulty of finding convenient class times, the need for travel to studios, the lack
of personalised coaching, the inability to track individual progress, the high cost of classes, and
the scarcity of customised content for specific health conditions like diabetes or heart disease.
Critically, incorrect posture during yoga can lead to adverse physical consequences, such as
muscle injury, joint pain, back pain, nerve compression, and even internal organ damage.
Therefore, there is a clear demand for solutions that enable learners to focus on proper alignment
and breathing, ensuring a safe and effective yoga practice [3]. To address these limitations,
an innovative technological approach is essential. Our primary objective is to develop an AI-
driven algorithm capable of providing real-time yoga pose correction. This algorithm will offer
immediate feedback to practitioners, leading to improved health outcomes and significantly
reducing the reliance on human instructors. Given the dynamic nature of yoga poses, the system
must be designed for real-time processing and maintain a lightweight architecture to ensure
seamless performance. Our innovative solution leverages the Convolutional Neural Network
(CNN) model, specifically MoveNet, for processing videos of yoga practice. This algorithm is
selected for its high accuracy and its ability to support multi-person scenarios. To facilitate its
training, a custom dataset was meticulously prepared, encompassing ten fundamental yoga
poses: Bhujangasana, Garudasana, Halasana, Natarajasana, Sukhasana, Makarasana, Savasana,
Simhasana, and Tadasana. Furthermore, we have incorporated insights from Yoga-82, a
comprehensive dataset designed for fine-grained classification of human poses, which offers
a three-level hierarchy encompassing body positions, variations, and specific pose names. This
technological approach forms the core of our "Virtual Yoga Teacher with AI Model for Yoga
Posture correction for Good health," a concept elaborated upon in our publication by Swapna
Yenishetti, Ganesh Karajkhede, and Lakshmi Panat. The integration of AI, particularly CNN
models, offers profound benefits for both preventive and curative healthcare. As a powerful
tool for real-time human pose estimation and tracking, AI significantly enhances the outreach
of yoga, transforming it into a non-invasive and sustainable healthcare model accessible to

19