We build computer-vision systems that watch match and training video, track every player and turn what they see into tactical insight, physical data and broadcast-ready graphics.
Twelve clips, from broadcast matches to training sessions — each one a computer-vision pipeline of ours running on real video.
Passing triangles, highlight rings and player spotlights drawn onto the match feed and anchored to tracked players, so they move with the play. An analyst names a player once and the tag follows them for the rest of the clip.
The court is calibrated from a single fixed camera, every player is tracked and assigned to a team, and speed, distance and sprint distance are measured from the video alone — no GPS vests.
Full-body keypoints for both players through fast rallies, reflections and overlaps — the raw material for footwork, movement and technique analysis.
Six players keep stable IDs through a drill full of crossings, runs and occlusions, so every position and action in the session is attributed to the right player.
Players are detected and split into teams by kit, and a team's shape is drawn as a network that moves with them — spacing, compactness and how the block shifts with the ball.
The court is detected in the broadcast picture and both players are projected onto a top-down map in real court coordinates, labelled by name — positioning and movement, point by point.
A single athlete is identified by name and followed through rotations, rallies and camera cuts in broadcast footage. Court positions, movement and involvement add up, rally by rally, into a profile of their playing style.
The ball — a few pixels across in a wide shot — is detected and tracked through passes and strikes, and its speed is read out in m/s as it travels. The same tracker measures the movement and speed of players or any other object in frame.
Six synchronized views of the same play, all computed from one broadcast feed: the tracked players, team labels, the network of connections between teammates, the area the team occupies and the lines its units hold.
Every point of the pitch is assigned to the nearest player, and the cells are coloured by team and redrawn every frame. The bar on top turns it into a number: the share of the pitch each side controls.
An athlete's skeleton is tracked through a footwork drill filmed with one ordinary camera, and every step is counted and the distance covered added up as they move — a running total on screen, with no sensors or wearables.
As the broadcast camera pans and zooms, both teams are projected onto a 2D pitch map, frame by frame. The shaded cone shows which part of the pitch the camera currently sees — so positions stay true even when half the field is out of shot.
We start from what you already record: a broadcast feed, one fixed wide camera, a drone or a phone at training.
Detection, tracking, pose and calibration models tuned to your sport, venue and camera angle.
Outputs are checked against hand-tagged footage before anything ships, so the numbers hold up in front of coaches.
Rendered video, raw data (CSV / JSON), an API or a web dashboard — whatever fits your workflow.
Led by Dmitry Kalinin, with more than twenty years of building and running production software. Today that experience goes into AI for sport — the models, and the products that put them in front of coaches, analysts and fans.
Labelling tool for sports video — turns match footage into annotated training data for our detection and tracking models.
Video-based statistics tracker — annotate match footage, log events by timestamp, export to CSV.
Football analytics calculator — real-time xG, xT, xA and xPass for any spot on the pitch.
The same models adapted for futsal, with save probability and zone metrics on a 40×20 m court.
Post-training check-ins and load monitoring — players answer in 30 seconds, coaches see team trends.
Send a clip and a question — we'll tell you what the models can get out of it.
dmitry@kalinin.technology