POSE ESTIMATION
FOR FITNESS WORKOUTS
THERAPY APPLICATION
Overview
Artificial recognition is a blessing for humanity as it has a lot to offer us in several areas of life. Moreover, it has played a significant role in the health and fitness sector. Artificial Intelligence-based mobile apps have created a buzz these days, especially during the pandemic when the gyms are not operating and people need someone to guide them with the workouts and postures. AI has performed an incredible job as it helps users work out perfectly well in real-time. We created an AI-based mobile app for our client to detect a human posture in real-time as their previous model didn’t do that well
Client Requirement
Our California-based client owns a startup and has created a mobile app to capture and evaluate human body motions during exercise and physical therapy. The mobile app is efficient with tracking tools to help users work out in the right way. However, the open-end solutions for posture estimation didn’t work well for our client. Hence, they approached HData Systems on the issue of human pose estimation in real-time. We developed a human pose estimation solution in their mobile app using AI that determines the human body orientation and position with a picture including a person.
Solutions​
- After understanding our client's requirements, we concluded that 2D single-person posture estimation in real-time would be an ideal option when implemented in different physical workouts.
- Moreover, we incorporated deep learning to detect human body movements in real-time. Our developers built a new neural network tech to use profound insights and ideas to achieve top-notch joints detection quality.
- We began with combining all open datasets for multiple types of human pose estimation as high volumes of data are the main ingredient of top-quality deep neural networks to perform with precision and strength in particular real-life use cases.
- To offer clients a seamless training pipeline, we used PyTorch, and for deploying CV models and deep learning, we used CoreML models.
- Our experts implemented data science algorithms for error detection during workouts. So when the user does an exercise incorrectly, an error is detected, and the app guides the users on how to do it the right way to avoid any injuries.
Challenges
- Build an ultra-modern posture estimation model to detect a human pose in real-time.
- Our main task was to enhance the accuracy of human posture estimation without affecting its speed and usability.
- Image attack
- Hence, our HData team has to create the data processes collecting particularly for that task and gather the data needed for training.
Technologies
Get In Touch With Us.
One World Trade Center, 285 Fulton Street suite 8500, New York, NY 10007, United States1
alibaba.softwaresolution
info@alibaba.com
