YoFlex (Yoga Trainer)

YoFlex empowers yogis, new & advanced, to achieve their full potential through a mobile, cost-effective, personalized yoga trainer experience that helps them improve form, track flexibility & minimize risk of injury.  Through deep learning technology and the latest computer vision techniques we provide the feedback you need to advance your practice and improve your quality of life.

Motivation

We've seen a global spike in yoga practitioners in recent years. Worldwide, approximately 300 million people regularly practice yoga, and approximately 36 million Americans practice yoga regularly. In the past 5 years alone, yoga practitioners in the U.S have increased by over 50%. 

While completing her yoga teacher certification, it became clear to one of our team members that students often do not obtain the appropriate level of feedback in group classes, and cannot obtain it through most mobile applications, AI powered equipment, and online classes. Advancing to certain poses too quickly, or using improper form, can lead to strain and injury. One on one sessions are often expensive, and not everyone sets out to gain a solid understanding of the fundamentals through yoga teacher training. Furthermore, methods to track progress are common in the fitness community, though lacking in the flexibility progression space.

YoFlex provides you with an understanding of the basics, offers options for modifications, gives you a baseline of your body in each pose, and helps you progress your form and flexibility in each pose over time.  On top of this, knowing the limits of your range of motion can help you prevent injury. We provide you with the training experience you need, at home, to help you excel safely in the studio and beyond.

How It Works

Once a specific yoga pose is detected, using a Resnet50 neural net trained with the Yoga-82 dataset, audio prompts the user better feel the shape and to correct their body in that pose. 3D pose estimation, which is also running from the moment a user logs into the app, extracts landmark key points on the user's joints and calculates major joint angles relevant to the detected pose.

The user is prompted with the captured image and data of themselves in the pose, a modified version of the pose (to provide an option to gradually advance in the pose), and an image of the "ideal" pose position (so they have a better understanding of the shape). The user's progress over time in the pose is displayed, along with a graph comparing their major body angles to the ideal angles for the pose. It is important to note that not everyone's body fits the "ideal" shape and YoFlex accounts for this.

After logging into the application and authenticating, the user must agree to a privacy policy and terms and conditions to use the application.  The video data is not saved, though an image of the user in the final version of the pose and their corresponding data around that pose is. On top of this, the user is provided with advice on how to prevent injury when logging into the application.  A series of warm-up poses is also recommended for the next time they attempt each position. 

We went beyond building models and focused on their holistic application. We are excited to continue expanding this technology at the intersection of the yoga, fitness app and mHealth markets.

Acknowledgements

We thank our capstone instructors Joyce Shen and Fred Nugen for thier feedback and guidance.  We also thank the creators of the Yoga-82 dataset for allowing us to use the dataset for non-commercial research and educational purposes.