Page 358 - Kaleidoscope Academic Conference Proceedings 2024

2024 ITU Kaleidoscope Academic Conference

1. The user interface is designed to be intuitive,
engaging, and visually appealing, ensuring ease of
navigation and accessibility for all users.
2. Features clear and concise information on
sustainable practices, making it easy for users to
adopt and incorporate these practices into their
daily routines.
3. The application leverages data analytics to provide
personalized recommendations and insights based
on user behavior and preferences.
4. Utilizes data-driven approaches to track and
measure the user's environmental and social
impact, providing feedback and rewards for
positive contributions.
5. Offers a wealth of educational resources such as
articles, videos, and interactive modules to raise Figure 1– Landmarks
awareness and educate users about sustainable
yoga practices.

2. DATASETS
5. OBJECTIVE AND METHODOLOGY
Yoga-82 is a new Dataset for Fine-grained Classification of Objective: Development of DNN based detection model
Human Poses. It is one of the largest data set which is used using transfer learning technique to detect the yoga pose,
to train the yoga classification techniques. The data set capture landmarks on real-time image or video to detect the
contains a three-level hierarchy including body positions, posture and integrate it in application to alert user for
variations in body positions, and the actual pose names. We correctness. The application shall also have encyclopedia on
present the classification accuracy of the state-of-the-art yogasana’s with relevant information.
convolutional neural network architectures on Yoga-82. We
also present several hierarchical variants of Dense Net in Methodology: For examination of different yoga poses,
order to utilize the hierarchical labels. On similar lines, we pre-trained model MoveNet is used, which detects 17
have also created custom data set to train the model for landmarks on each images and creates csv file and prepared
most common poses used by the public to demonstrate the train and test data. These manually created csv files are fed
capability of the model.
to train the model for classification of poses. We have
developed react.js front-end application and used
3. LOGIC
tensorflow.js for importing the model. At first it shall detect
the pose and capture 17 landmarks from input image or
Human pose estimation is a Computer Vision technique video for comparison with master landmarks. If current
used to predict a person’s body parts or joints position. This posture reaches the accuracy of threefold above 70%
can be done by defining the human body joints like wrist, probability then, the colour turns to the green and also it
shoulder, knees, eyes, ears, ankles, arms, also called key counts the time for holding the posture correctly and
points in images and videos. Then, when a picture or video notifies us by playing sound.
comes in as input to the pose estimator model, it identifies
the coordinates of those detected body parts as output and a 6. LITERATURE SURVEY
confidence score indicating continuity of the estimations.
The MoveNet model is based on 3D estimation. The Due to the increase of stress in the modern lifestyle, yoga
operation takes place in a phase-wise manner like; first, the has become popular throughout the world. Although
RGB image is fed to the convolutional network as input, considered to be a great way to bring physical, mental, and
then the pose model is applied to detect the poses, key spiritual harmony, it can come with its own set of problems,
points, pose confidence score and key point confidence if not done properly. This gave rise to the need for Yoga
score from the model outputs.
Pose Detection and Classification, to omit the necessity of a
human instructor. But, for an AI agent to behave like a
4. LANDMARKS
human, it is necessary to have a strong model to detect and
classify various yoga poses. OpenPose was the pioneer
The landmarks of a human body used in this project are project in the field of Landmark detection using Image
nose, ears, eyes, shoulders, elbows, wrists, hips, knees, and Processing. Using 18 landmarks detected on the human
ankles. This gives us 17 landmarks, which will be used to body, it was able to achieve 78% accuracy in detecting the
detect various poses by the model. yoga poses. PifPaf uses a Part Intensity Field (PIF) and a
Part Association Field (PAF) for body part limitation, and
relationship of body parts to shape full human stances,
respectively. This technique is dependent on the base up

– 314 –