Spatiotemporal Recognition of Figure Skating Jumps Using 3D Convolutional Neural Networks

A deep learning-based system for recognizing figure skating jumps in video recordings is investigated in this paper. It is particularly hard to identify different jump types including Axel, Salchow, Toe Loop, Loop, Flip, and Lutz accurately due to the brief duration and subtle technical variations among them. To address this obstacle, footage from Olympic competitions was segmented into short, labeled clips, forming a customized dataset. To learn spatiotemporal features of each jump type above, a three-dimensional convolutional neural network (3D CNN) was trained on these temporally aligned segments. Results from experiments indicate that, compared to analyzing entire jump sequences, classification accuracy can be enhanced by focusing exclusively on the take-off phase. The trained model is integrated into an interactive web platform so that users could upload videos and receive real-time feedback on jump classification. The practical applicability is also suggested in judging support systems and athlete training.