Classification of skateboarding tricks by synthesizing transfer learning models and machine learning classifiers using different input signal transformations

• Main Author: Muhammad Amirul, Abdullah
• Format: Thesis
• Language: English
• Published: 2022
• Subjects: TJ Mechanical engineering and machinery; TS Manufactures
• Online Access: http://umpir.ump.edu.my/id/eprint/35919/1/Classification%20of%20skateboarding%20tricks%20by%20synthesizing%20transfer%20learning%20models%20and%20machine%20learning%20classifiers.wm.pdf; http://umpir.ump.edu.my/id/eprint/35919/

Skateboarding has made its Olympic debut at the delayed Tokyo 2020 Olympic Games. Conventionally, in the competition scene, the scoring of the game is done manually and subjectively by the judges through the observation of the trick executions. Nevertheless, the complexity of the manoeuvres executed has caused difficulties in its scoring that is obviously prone to human error and bias. Therefore, the aim of this study is to classify five skateboarding flat ground tricks which are Ollie, Kickflip, Shove-it, Nollie and Frontside 180. This is achieved by using three optimized machine learning models of k-Nearest Neighbor (kNN), Random Forest (RF), and Support Vector Machine (SVM) from features extracted via eighteen transfer learning models. Six amateur skaters performed five tricks on a customized ORY skateboard. The raw data from the inertial measurement unit (IMU) embedded on the developed device attached to the skateboarding were extracted. It is worth noting that four types of input images were transformed via Fast Fourier Transform (FFT), Continuous Wavelet Transform (CWT), Discrete Wavelet Transform (DWT) and synthesized raw image (RAW) from the IMU-based signals obtained. The optimized form of the classifiers was obtained by performing GridSearch optimization technique on the training dataset with 3-folds cross-validation on a data split of 4:1:1 ratio for training, validation and testing, respectively from 150 transformed images. It was shown that the CWT and RAW images used in the MobileNet transfer learning model coupled with the optimized SVM and RF classifiers exhibited a test accuracy of 100%. In order to identify the best possible method for the pipelines, computational time was used to evaluate the various models. It was concluded that the RAW-MobileNet-optimized-RF approach was the most effective one, with a computational time of 24.796875 seconds. The results of the study revealed that the proposed approach could improve the classification of skateboarding tricks.