Our client runs a Yoga School and wanted to take his skills to everyone that wanted to practice the ancient art for healthy living.
The business asked for an AI Yoga trainer that could evaluate the user's body asana posture to determine if they were doing the yoga pose correctly.
An accurate Yoga pose detector and assessor that offers inputs on what to change to get a more accurate pose. Additionally, offering comparison charts of users to show progress.
To begin, we needed to create a system that could fetch feed from the camera.
After appropriate Research and Development, we opted for the OpenCV HAAR CASCADE algorithm as the means to detect faces.
Additionally, the detection system is aligned with the task of recognizing the subscriber through image analysis to enable access to the web application.
Lastly, we custom-trained our model with CNN of LSTM to classify Yoga Poses.
Start by turning on the web application and give access to the camera. Then, the application will start capturing the feed.
Afterward, the current feed will get sent to the server where our AI model is deployed.
In return, the AI model is responsible for offering feed that gives the pose category and angle of every joint.
Here, it will assess all the angles of your joint, and another ML algorithm will classify whether the pose is correct or not. It will provide insights to ensure that the user can perform the activity accurately without injuries.
With the help of a face recognition algorithm, the application only allows registered users to use the application.
The web application stores data related to your previous activities on the app. Then, it can offer insights on improvements you are making in the exercise, giving it a more human feel while training.
Depending on the user's location and internet connection health, low bandwidth is a possibility. In such a scenario, the image capturing was inadequate, which led to poor results.
We optimized the algorithm to change the image resolution as per the bandwidth of the product, enabling the product to work even when working with low bandwidth.
We were tasked with training the model on a limited number of images, and that led to issues with results when the angles or placement of the camera changed.
To overcome the issue, we used one-shot learning.
By the end of the project, we created a smart AI-powered yoga trainer that can help people live healthier lives and get expert advice while they do yoga at home. The AI-powered web application uses joint angles to determine the quality of yoga poses and provides helpful insights to the users to help them perform these poses accurately and avoid injuries.