Motion Based Trick Classification in Skateboarding Using Machine Learning

The success of sports related applications such as Strava and Runtastic showed the potential for personal progress tracking and competitive digital leader boards. All of the applications rely on some hardware, that records relevant data for the respective platform in an automized fashion. Communities for running or cycling have benefited from an abundance of such systems whereas the skateboarding scene only has limited options. Therefore, the goal was to design an create a proof of concept for an automated trick detection in skateboarding. For an automated trick evaluation, a neural network was trained on a dataset containing 1300 motion data samples of 5 different skateboarding tricks. The motion data of twelve athletes was recorded with an inertial measurement unit attached to the front truck of the skateboard. The location of the sensor on the truck was deliberately chosen in a way so that it does not interfere with the athletes normal routine at any time during skateboarding. By using data augmentation, synthetic samples were generated to ensure an equal number of samples for each trick class in the dataset. The task of the network is to classify tricks based on their characteristic motion data represented as a time series. A convolutional neural network (CNN) was created using the PyTorch framework. With an accuracy score of 95,1% the network is capable to correctly classify the majority of tricks and therefore proving the proof that a trick classification in skateboarding is possible with the given concept.