Skateboarding Trick Classification Using Transfer Learning-Based Image Processing and Support Vector Machine

This study presents a novel approach to classify five skateboarding tricks (Kickflip, Frontside-180, Nollie Frontside Shove-it, Pop Shove-it, and Ollie) using transfer learning models integrated with Support Vector Machine (SVM) classification. As skateboarding continues to gain prominence in competitive sports, including its Olympic debut, there is increasing demand for objective evaluation systems. The methodology captures skateboarding trick sequences using a YI action camera positioned 1.26m from the performance area and extracts image frames at 30fps. By overall of approximately 750 images were extracted and then would proceed through a train, validation, and test split of 60:20:20 ratio, respectively. Four pre-trained CNN architectures (NasNetLarge, NasNetMobile, MobileNetV2, and MobileNet) were evaluated as feature extractors coupled with SVM classification. Comprehensive evaluation revealed that NasNetLarge achieved the highest classification accuracy of 93% on the test dataset, followed by NasNetMobile (92%), MobileNetV2 (91%), and MobileNet (87%). Confusion matrices indicate specific patterns of misclassification between similar tricks. This objective evaluation system provides a foundation for automated trick recognition in competitive skateboarding, offering potential applications for objective judging systems in competitions and as a training tool for skateboarders seeking performance improvement.