Teenage tech stories
Each month, Tech for Good speaks to one teenage entrepreneur about their incredible achievements in the world of tech, and how they’re contributing to making the world a better place.
Name: Michelle Hua
Age: 16
Born: Michigan, USA
Achievements: As a coder and competitive gymnast, there’s little that Michelle can’t do. The 16-year-old has won the $75,000 George D. Yancopoulos Innovator Award for its movement-tracking algorithm, that she developed during lockdown to help her with her gymnastics training. Her research for this project was later published in the journal Computer Aided Geometric Design.
I started doing gymnastics when I was really young, around seven or eight years old. My parents wanted me to do something outside of school to keep me active so they gave me the choice between more popular sports, like soccer and basketball, but I chose gymnastics, and I’ve been doing that ever since.
I was inspired to do my project due to the COVID-19 pandemic. Since we were in lockdown and we had to stay at home, people had limited opportunities to go out to gyms to exercise, including myself, because my gymnastics training was moved online.
I wanted to create an algorithm to recognise movements and then give feedback, which can be used in our coaching, and also in a lot of different real-world applications including physical therapy, sport analysis, autonomous driving and public safety.
Basically, you can give my algorithm a video of anyone doing an action, and it will first extract the silhouette’s boundaries, so like a line around the human for all different frames. Then, my algorithm uses dilated point convolutions to find patterns and features in action, and through those features and patterns, it can recognise what category the action belongs to.
I use silhouettes, which is a novel representation that no one’s ever been using in deep learning for action recognition before. They normally use skeletons. And those are actually less robust than mine, because they require a lot of knowledge about the location of joints, for example, where your arm is located, where your legs are located, and its orientation. But for silhouettes, all I have to do is separate the human from the background it’s in. And then through this accurate representation, I’m able to more accurately recognise action.
The hardest part was probably designing my dilated point convolutions, because normally for things like AI and deep learning, people just use regular 2D convolutions. It’s like a really traditional method to extract features. But then since my silhouettes are stacked into a 3D-point cloud, it’s in 3D. So I have to use a novel 3D method to extract these features and patterns, instead of normal 2D methods.
To train my algorithm, I used three benchmark datasets. They’re just collections of different videos of different actions, and it can all be found online and publicly available.
Some types of movements weren’t available in the dataset, like my gymnastics moves or some physical therapy moves, so I had to create my own data. I filmed myself doing those movements and labeled the video with the action that is occurring, so I could put it in my algorithm and train it.
My project is focused on my algorithm, because that’s where I spent most of my time. And the app is a real-world application or a practical need that I applied my algorithm to.
It can definitely be used for a lot of different things like sports training, and general exercises, just to keep yourself healthy and active, especially during the pandemic. And I also included some physical therapy exercises, so if you can’t go to any physical therapy clinics, you can practice at home and receive real-time feedback about how well you do the exercises.
Right now, the app detects around 20 actions. In the next few months, I hope to keep adding different exercises and different sports that you can practice. And I’m actually thinking of submitting my app to the Apple store so that more people can download it and use it.
In my science and math classes, there tend to be more males. But I think it hasn’t discouraged me because I know that I work hard and they work hard too; so we’re all on the same page. And for the research paper, two of the people that presented in the conference with me were females, so I think things are getting better.
It was my first time publishing a paper, so it was definitely a lot of work. Even after writing the paper, they would come back with comments, and they’ll have to revise it over and over. But it was very exciting.
I want to continue doing research, specifically in AI and deep learning. I think AI can help out and influence and benefit a lot of different fields of science. And I’ll definitely continue throughout high school and college doing competitive gymnastics. Last month, I actually competed in the USA National Gymnastics Championships, and I want to go again next year.
