An automatic hockey camera that runs on one desktop per rink

Each stage closed with a written report and a working build, and each started from what the previous report showed.

1. First prototype. Player detection, masking of the rink and a first region-of-play tracker, tested on the client’s own game footage.
2. Setup tools the client runs on its own. The prototype showed where false detections came from, so we built a GUI to draw the rink mask and set up stitching. From then on, the client’s team could open a new rink without us.
3. Speed. We profiled the whole chain on a live stream, then built a new tracker and camera controller around what the profiling showed.
4. Panorama v2. A stitcher for two cameras with fewer manual steps and a choice of projections.

Why it matters: the camera can only follow the play if it knows where every player is, in real time, on ordinary hardware. If it reacts to people on the bench or in the stands, viewers watch the crowd instead of the game.

Before looking for players, the system masks out everything that isn’t the playing surface. The mask is drawn by hand in the setup tool or found automatically from the colour of the ice, once per rink.

Rink lighting stays the same through a game, so the empty ice works as a stable background. The detector compares each frame with it and keeps what changed. Players are darker than the ice, so a one-sided comparison is both faster and cleaner. Shapes are then filtered by size and fill, scaled by their place on the ice, because a far player looks smaller than a near one.

Why it matters: false detections make the camera twitch, and lost players make it miss the play. This is the step that frames the game for the viewer.

Each player becomes a track. A new track counts only after it is confirmed on 30 frames in a row, and a track is dropped after 10 frames with no match, so a passing shadow never moves the camera. An alpha-beta filter smooths each track and predicts where the player goes next. When players pass close to each other, tracks are matched by how much their boxes overlap, strongest match first.

The system then groups players into clusters and weighs each one. A small cluster far from the action is dropped when that makes the frame much tighter, so one player skating to the bench doesn’t widen the shot. The frame keeps its aspect ratio, stays within the zoom limits and never leaves the rink.

A camera controller smooths the frame over time and predicts its next position, so the view pans like an operator’s and holds steady through a quick pass.

Why it matters: two cameras joined badly show a seam exactly where the play crosses.

The stitcher joins two wide-angle cameras into one panorama of up to 180°. Version 1 needed a calibration and several manual steps per camera. Version 2 needs only 4 to 10 matching point pairs, set in a GUI, and computes the rotation between the cameras automatically. Spherical, cylindrical and flat projections can be compared on the spot. The setup is done once per rink.

Why it matters: a tracker that keeps up on a test file can still fall behind on a real game, and viewers notice a stall at once.

We profiled the whole application on a live game stream: CPU-only and with CUDA, with software decoding and with the NVIDIA hardware decoder. One finding shaped the rest: on large frames, copying images between the CPU and the GPU took more GPU time than the image processing itself.

We rebuilt the tracker and the camera controller around what we found, moved colour conversion to a faster library, and made every stage configurable with a debug view. The application reads one live stream and sends out two: the panorama and the tracked camera view.

A live stream leaves about 30 ms per frame for everything. We measured the tracker, the full chain and the load on a desktop.

A business like this grows by the number of rinks it covers. With the framing automatic and the setup in the client’s hands, a new rink takes a camera, a desktop and one setup.

• Growth by rinks, not by staff. Every new rink gets a camera and a one-time setup, and its games are framed with no one filming.
• Games worth rewatching. A view that follows the play reads like TV coverage, which a fixed wide shot of the whole rink never does.
• No special hardware. The tracker keeps up on an older laptop CPU, and a live game runs on one mid-range desktop.
• No waiting on outside engineers. The client’s own team sets up every new rink in the GUI.