Watch Me

When working out, recording even simple workout information can be tedious and inconvenient. And for individuals to assess their own form is often impractical at best. Our app Watch Me, can automatically track and monitor your form down to each individual repetition.

Our design hinges on several emerging and rapidly developing technologies, specifically AI/Machine learning. Using machine learning algorithms, we were able to detect aspects of upper body exercises using the accelerometer and gyroscope data taken from right on your smartwatch.

Our product aims to provide users a way to use their wearable devices to provide real-time qualitative feedback on their exercise form at a cheap price. This feedback will reduce the risk of injury and improve the benefits received from performing the exercise. Compared to traditional solutions of form assessment, such as a personal trainer, our solution can save customers thousands of dollars a year.

We validated our app design by going through a round of UI/UX testing and research. In this testing, we focused on making our app usable, enjoyable, and equitable. We sent surveys out to individuals and measured statistics such as task completion rate and customer satisfaction. We validated our machine learning algorithm by using many of the standard assessment methods. These include classification accuracy, logarithmic loss, and confusion matrices

Through market research, the team has prepared financial forecasting to determine if the product is feasible enough to generate profits. We found that initial startup cost would begin at roughly $100,000 and would require even more support for the second and third years of the startup’s life. After year four is when we first start seeing profits, increasing each year past this point. For this reason, the team members agree that the startup would not be financially feasible on our own, and as an alternative we would be looking to sell our idea to a larger company with suitable capital. 

The smartphone application uses a gaussian windowing technique to isolate the start and stop of each individual repetition and a K nearest neighbor algorithm to categorize the form. The current implementation has repetition detection and form evaluation accuracies of 90 and 84 percent respectively. These metrics were generated from the initial MVP implementations of the algorithms which can be greatly improved. Given the growing competition in this market and our current successes we believe this solution is realistically feasible.