The Classification of Skateboarding Tricks: A Support Vector Machine Hyperparameter Evaluation Optimisation

The growing interest in skateboarding as a competitive sport requires new motion analysis approaches and innovative ways to portray athletes’ results as the conventional technique of the classification of the tricks is often inadequate in providing accurate and often biased evaluation during competition. This paper aims to identify the suitable hyperparameters of a Support Vector Machine (SVM) classifier in classifying five different skateboarding tricks (Ollie, Kickflip, Frontside 180, Pop Shove-it, and Nollie Frontside Shove-it) based on frequency-domain features extracted from Inertial Measurement Unit (IMU). An amateur skateboarder with the age of 23 years old performed five different skateboard tricks and repeated for five times. The signals obtained then were converted from time-domain to frequency-domain through Fast Fourier Transform (FFT), and a number of features (mean, kurtosis, skewness, standard deviation, root mean square and peak-to-peak corresponding to x–y–z axis of IMU reading) were extracted from the frequency dataset. Different hyperparameters of the SVM model were optimised via grid search sweep. It was found that a sigmoid kernel with 0.01 of gamma and regularisation, C value of 10 were found to be the optimum hyperparameters as it could attain a classification accuracy of 100%. The present findings imply that the proposed approach can well identify the tricks to assist the judges in providing a more objective-based evaluation.