Technology
Hockey is one of the harder video domains: boards, glass reflections, thirty-six bodies on a small sheet of ice and a puck that is eight pixels wide.
A single-stage detector runs on every frame; a multi-object tracker maintains identity across occlusion, with re-identification when players leave the frame.
Seventeen keypoints per player per frame, refined with temporal smoothing. Used for skating mechanics, hit detection and injury review.
A dedicated small-object model plus trajectory fitting. When the puck is occluded we predict from the last confident observation and flag the gap.
Ice markings are detected to recover the homography, so coordinates are in metres on the rink rather than pixels on a frame.
Zone entries, exits, shots and passes are classified from trajectory and pose features, then reconciled with the official scoresheet.
All layers share one clock. A clip requested at 14:32.400 returns those frames, not the ones twelve milliseconds later.
Video analytics is an infrastructure business. These are the numbers behind a " normal month across the customer base.
| Layer | Monthly volume | Design target |
|---|---|---|
| Ingest | ~74 TB | Resumable, checksum-verified |
| Transcoding | 320,000 hours | Six renditions per source |
| GPU analysis | ~48,000 GPU-hours | Real time or faster per feed |
| Storage | 4.2 PB | Hot 14 days, cold thereafter |
| Delivery | ~1.4 PB | CDN, 40+ points of presence |
Reliability
A game starts when it starts. Ingest and playback run on separate paths so a saturated processing queue never stops a coach from watching a clip that already exists.