PitchIQ

PitchIQ turns video of a high-school football (soccer) match into stats a coach can use: possession, territory, shape, a top-down map of every player, shots and xG. Players get reports of their own as well. A tablet on the touchline tags what the camera can't settle by itself (kick-off, substitutions, restarts). The computer-vision pipeline in pitchiq/ turns the match video and that tag log into a report JSON, and the site (GitHub Pages and Firestore) shows the report to the coach and the players.

PitchIQ is source-available under the PolyForm Noncommercial License 1.0.0: free for non-commercial use, see License.

Status

Working on the team's own footage: a panning (Hudl) camera calibrated from the match itself, the Hybrid 90 ball tracker, and reports on the site for the coach and the players. It is a research tool used by one team, not a finished product. The ball tracker finds about three balls in four on held-out clips and abstains on the rest, by design, but under floodlights it loses the ball almost entirely. Players and ball together on one top-down map is the next piece, and real footage needs a CUDA GPU. ROADMAP.md has where things stand and what is next.

Quickstart

Python 3.12, from a clone of the repo (Setup has the rest):

pip install -e .[test]
python -m pitchiq demo

demo draws a nine-second scripted match (coloured boxes: a long pass, a touch, a shot), fits a fixed-camera calibration to it and runs analyse_match with a scripted detector standing in for YOLO. It needs no footage, GPU or weights and finishes in seconds; --out DIR keeps the clip, the calibration and the report JSON. The numbers show the chain working, not how well it sees a real match. On real fixed-camera footage the same steps are grab_frame, calibrate and match_report (see FOOTAGE_DAY.md).

Two camera pipelines

Fixed camera (pitchiq/calibration.py). One homography per camera position maps pixels to metres. It is fitted once from landmarks clicked on a single frame. This is the pipeline FOOTAGE_DAY.md was first written for. A tripod, a locked-off shot, and that is all it needs.

Panning camera (pitchiq/panning_camera.py + pitchiq/balltrack/). The footage the team actually has is from a Hudl camera that follows the play. Every frame gets its own frame-to-pitch homography:

• pitchiq/autocal.py calibrates the camera from the match itself, using the paint plus SIFT between frames, and
• pitchiq/balltrack/pitch_registration/register_frames.py registers frames to a venue map.

The ball comes from the Hybrid 90 tracker in pitchiq/balltrack/, and PannedCamera moves everything into one still reference view. After that, pitchiq/pipeline.py runs unchanged. This example requires a GPU and generated data, so it does not run from a fresh clone:

from pitchiq.panning_camera import PannedCamera
from pitchiq.pipeline import analyse_match

camera = PannedCamera.from_tracker('fr00')   # needs hybrid.sh outputs, see below
report = analyse_match(video, panned=camera, start_s=..., end_s=...)

from_tracker reads files that sh pitchiq/balltrack/hybrid.sh <lo> <hi> <tag> (one window) or pitchiq/balltrack/half.py (a whole half) writes. They are generated outputs, not in git (see pitchiq/balltrack/README.md, "Generated, not in git"). Accuracy figures for the tracker are in that README too, and FOOTAGE_DAY.md ("Fixed or panning") sets out which pipeline a camera needs.

Repo map

path	what
pitchiq/	the pipeline: detection, tracking, identity, calibration, ball, possession, events, top-down map, report JSON, publishing to Firestore
pitchiq/cli/	the match-day commands, run as python -m pitchiq <command>: grab_frame, calibrate, event_report, match_report, compare_reports, publish_report, demo
pitchiq/experiments/	research tools and labs: spike_detect, track_report, speed_report, autocal_run, player_lab, ...
pitchiq/balltrack/	the ball tracker for panning footage (Hybrid 90: detector + classifier + TAPNext++ bridges + 3D flight chain), and the venue map and registration in pitch_registration/
shared/	shared frontend modules and CSS (no build step, plain ES modules)
index.html	landing page and sign-in
coach/	coach dashboard: season, roster, matches, staff, player reports
player/	a player's own reports
halftime/	the three-minute read for the touchline
live-tagging/	match-day tablet tool
calibrate/	landmark pins on a frame or a panorama, exported for both calibrators
xg_training/	xG model training (main.py, StatsBomb open data -> ONNX)
baselines/	committed report JSONs that compare_reports checks changes against
tests/	Python (pytest) and JavaScript (node --test, Firestore emulator) suites
docs/	architecture, glossary, report schema, conventions, per-module design notes, history and generated figures
tools/	repo maintenance (audit_metrics.py)

Setup

Python 3.12 (requires-python = ">=3.12"; CI tests 3.12 only). Clone the repo and install it editable into a virtual environment. That is the only supported install: the ball tracker and the panning-camera path read data files that sit beside the code (the venue map, the ball ranker and classifier), and a wheel or a plain pip install . does not carry them.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗