04Solutions family

Physics and features

Distances, arrival times, pitch control and filtered trajectories. Label-free evidence that can be read to the last number.

Pitch control, the best-known member: the pitch divided by who would reach each point first.
  • 6entries
  • 0need labels
  • 0trained weights to ship

One picture: positions become evidence

Tracking gives positions. Possession evidence needs more: who can reach the ball, how fast it is moving, whether its motion has just changed. This family computes those quantities from physics and geometry. Nothing is trained, so nothing needs labels, and every number traces back to a speed, a distance or a time.

The picture above shows the computation at full scale, over the whole pitch. For possession the same computation is needed at one point only: the ball.

The standing receiver is closer. The sprinting presser arrives first. Time to intercept gets right what distance gets wrong.

In the chain, measuring comes after the data is read and aligned, and before anything decides. Its output is evidence for the last link.

How to choose inside it

Entry Labels Ball and identities Compute
Pitch control None Ball position for local use. No identities Light at the ball, heavy on a full grid
Accessible space None Ball required. Some functions want the team on the ball Heavier: simulated passes
Kalman filter and RTS smoother None Ball measurements, gaps allowed Negligible
ruptures None Any signal: ball speed or team speed Light
floodlight None Whatever is tracked Light
unravelsports None for features One node per player. No identities Light for features

Start with distances and speeds. Add time to intercept from pitch control where distance misjudges a sprinting player. Add change-points on ball speed where touches matter.

Raw ball speed places the touch to the frame and adds false ones. Smoothing removes the false ones and blurs the real one into a band. Every filter in this family sits somewhere between the two.

What this family cannot do

  • It cannot name the possessor. Controlling space is not having the ball. Being closest is not being in control.
  • It cannot make evidence where the sensor made none. A filter that bridges a long gap draws a plausible path, not a measured one.
  • It cannot be smooth and sharp at once. Every smoothing choice blurs the touch it was meant to reveal.

Where our decoder stands

Its evidence is this family’s kind of quantity, kept simple on purpose: a player within 1.5 m of the ball and closer than any opponent, the ball out of play, collective stillness of the players. A frame with no ball contributes no ball evidence, and nothing is imputed. The physics stays evidence. The Viterbi decode makes the decision.

In this family6

  • Who would reach each point of the pitch first. The physical model behind most spatial football analytics, in its reference tutorial code and a fast implementation.

    • Tracking
    • Physics
    • No training
    • MIT
  • A physical pass model that asks where the team on the ball can actually play to, and how dangerous that space is.

    • Tracking
    • Physics
    • Needs possession
    • MIT
  • The classical estimator for a noisy, intermittently observed trajectory. It gives every ball position an uncertainty, and the uncertainty is the part possession inference needs.

    • Tracking
    • Classical
    • No training
    • BSD
  • Offline change-point detection. Give it a signal and it returns the instants where the signal's behaviour changes: candidate touches, restarts and stoppages.

    • Time series
    • Classical
    • No training
    • BSD-2-Clause
  • A typed data model for team-sport tracking: positions, events, pitch and per-frame codes as first-class objects, with kinematics and spatial analysis on top.

    • Tracking + events
    • Library
    • No training
    • MIT
  • Turns tracking into graphs, pressing intensity and formations. The shortest path from a kloppy dataset to a graph neural network.

    • Tracking
    • Graph
    • Library
    • MPL-2.0