Recently, Exercise Analysis has gained strong interest in the sport industry including athletes and coaches to understand and improve performance, as well as preventing injuries caused by incorrect exercise form. This work describes a system, USquat, that utilizes computer vision and machine learning for understanding and analyzing of a particular exercise, squatting, as proof of concept for the creation of a detective and corrective exercise assistant. Squatting was chosen as it is a complicated form of exercise and is often mis performed. For USquat, a Recurrent Neural Network is designed using Convolutional Neural Networks and Long Term Short Term networks. A sizable video library dataset containing numerous “bad” forms of squatting was created and curated to train the USquat system. When evaluated on test data, USquat achieved 90% accuracy on average. On a developed Android application that uses the resulting model, upon detection of “bad” squatting forms, it offers an instructive “good” video related specifically to the user’s bad form. Results including live application outcomes are demonstrated as well as challenging video results, problems, and areas of future work. Early work on the creation of a follow-on system to USquat that automatically generates a custom video of the user performing a correct action for the purpose of learning proper activity performance. Additionally, early work on a different version of USquat that explores an attention mechanism network is discussed.
|