Skating and acceleration analysis from video, with AI
A system that estimates player movement from video, measures skating and acceleration, automatically turns footage into clips, and gives coaches data they did not have before.
On-ice performance that was hard to measure.
Coaches could see on video that a player accelerated or braked, but had no numbers for it. Reviewing footage by hand was slow and did not scale.
Let computer vision do the measuring.
We built a pipeline that estimates a player's pose from video, computes skating and acceleration metrics and produces clips automatically. The analytics run in two phases: pose estimation storing keypoints and annotated video, then a rule-based pass computing findings, screenshots and clips. The original desktop recorder for two IP cameras was replaced by a cross-platform Tauri 2, React 19 and Rust application with ffmpeg and go2rtc, which records precisely defined thirty-second dual-camera clips; processing runs on a GPU server via Docker Compose. The coach gets clear outputs, not hours of raw footage.
Measurable data out of video.
The analytics are covered by 602 passing tests, a complete run against the production database with measured timings is recorded in the repository, and the work spans 807 commits across six repositories.