Figure Skating AI

A sports-tech field that uses computer vision and machine learning to analyze figure skating jumps and movements, supporting judging and training.

Figure Skating AI

Overview

Figure Skating AI is a technical field that uses computer vision and machine learning to quantitatively measure and analyze figure skating movements and to support judging and training. It aims to reduce judging controversies and provide objective feedback to athletes and coaches by calculating frame by frame such elements as jump rotation count, landing angle, jump height, and flight time, which are difficult to determine with the naked eye. In the 2020s, as the International Skating Union (ISU), national federations, and sports-tech companies have pilot-introduced or researched related systems, it has established itself as a representative application of sports engineering.

Main Content

Background: Judging Reliability Issues

Figure skating uses a structure in which multiple judges sum the technical elements score (TES) and artistic score (PCS). In particular, judging of jump underrotation (short of a quarter turn) has been at the center of controversy. If the rotation is short by less than a quarter turn, q is applied; if short by half a turn or more, <; and if short by one turn or more, << (downgrade). These distinctions can vary depending on the angle of the replay video and the judge's viewpoint. Judging controversies repeated at the Olympics and World Championships increased demand for automated analysis software beyond hardware such as super slow-motion cameras.

Technical Composition

Representative systems consist of the following pipeline.

Multi-camera calibration and 3D reconstruction: Aligns images from cameras at different positions to reconstruct the skater's joint coordinates in three dimensions. It can correct for the rotation axis and landing angle that are occluded from a single viewpoint.
Pose estimation and blade detection: Models in the OpenPose, MediaPipe, and HRNet families estimate body joints, while a separate detector locates the position and orientation of the skate blade to determine the moment of contact with the ice.
Event segmentation: Automatically segments the takeoff–flight–landing phases in continuous video, counts rotations, and calculates the relative angles of the shoulders, pelvis, and blade at landing.
Classification and evaluation models: Trains skeleton time-series data with CNNs, graph neural networks (GCNs), and transformers to classify jump types (Axel, Lutz, Flip, Salchow, Toe loop, Loop) and failure types.
Reporting dashboard: Visualizes per-jump height, rotation speed, takeoff speed, landing impact, and other metrics for use by coaches and athletes.

Judging Support Methods

The ISU has operated a judge support system with technical partners since 2019 and is considering ways to increase automation by combining super slow-motion replay with jump rotation analysis. However, the principle is to maintain a 'Human-in-the-loop' structure in which human judges still assign final scores; AI is defined not as a replacement for judging but as an auxiliary tool that provides supporting evidence. At international competitions, the first approach being applied is to automatically mark candidate segments that judges need to check.

Training and Analysis Applications

The area that spread faster than competition judging is the training field. By combining inertial measurement unit (IMU) wearables, pressure mats, and multi-camera analysis rooms, flight time and rotation speed of jumps are measured continuously, and landing postures with high injury risk are detected early. In Korea as well, motion-capture-based analysis support is being provided centering on the Korea Skating Union (대한빙상경기연맹), sports science research institutes, and university laboratories, and low-cost applications have also appeared that let users upload video filmed with a single smartphone and receive feedback on rotation count and landing angle.

Limitations and Challenges

Difficulty in quantifying artistry: Many point out that PCS areas such as musical interpretation, expressiveness, and program composition are difficult to quantify and unsuitable for AI evaluation.
Data bias and generalization: Models trained on specific athletes or specific rinks may lose accuracy when lighting, costumes, or camera angles change.
Regulations and responsibility: Who bears responsibility when a misjudgment occurs, the scope of disclosure of judging data, and personal information protection remain unresolved issues.
Cost and standardization: There are no international standards for camera placement and data formats, making comparison between systems difficult.

Latest Trends

In 2024–2025, there is a continuing move to expand the pilot scope of automatic jump rotation detection to major competitions such as the Grand Prix Series and World Championships, and AR telestration cases are increasing in which judging data is linked to broadcast graphics to show rotation count and flight time to viewers in real time. Research using pose foundation models and generative AI to reconstruct 3D skeletons even with a small number of cameras is active, and attempts are underway to manage athletes' cumulative season data as digital twins for conditioning and injury prevention. It is also noteworthy that, ahead of the 2026 Milano-Cortina Winter Olympics, national federations are allocating budgets for building analysis infrastructure. However, cautious views are dominant among both athletes and judges regarding the issue of handing final judging authority to AI, and for the time being the dual structure of 'judging assistance + training analysis' is expected to be maintained.