Current judging and scoring approaches at figure skating competitions requirehuman visualization of skills by a technical specialist in order to render a score. Factors such as visibility, time constraints, and fatigue after a long competition can impact a technical specialist`s ability to accurately classify figure skating jump attempts, thus introducing unnecessary bias in scoring. US Figure Skating Association has a pressing need and interest in automating their scoring methods to eliminate such biases. Therefore, I propose a video classification algorithm built using a temporal segment network (TSN) to classify the six jump types in singles figure skating: axel, salchow, toe loop, loop, flip, and lutz. By utilizing pose estimation on a set of 2.0 second video clips illustrating jump attempts from figure skating competitors, I optimized a TSN model constructed with a ResNet-50 backbone to predict these six classes with 62.22% mean class accuracy. The model does exceptionally well at predicting jumps with forward take offs (i.e.,axels) with an F1 score of 0.939; however, there is still room for improvement in classifying "edge jumps" (i.e., salchows and loops). In conclusion, this project lays the foundations for future figure skating classification algorithms to eventually distinguish inaccuracies of jump attempts (i.e., under-rotations or edge changes), thus establishing the frame-work for an automated technical specialist.
© Copyright 2021 Alle Rechte vorbehalten.