Analysis
Football tracking data explained: how the pipeline works
How football tracking data is captured by camera arrays and wearable pods, joined to event data, and converted into physical metrics and offside calls.
By CricketTaken EditorialPublished Analysis18 min read
A rig of cameras is bolted to the roof gantry above the top tier, angled down at the grass, and for the whole ninety minutes nobody in the ground looks at it once. It generates more raw information than every other source in the stadium put together, including the broadcast feed. Football tracking data explained properly has to begin there, with the hardware, because most arguments about what the numbers mean turn out on inspection to be arguments about how they were captured.
Strip away the vocabulary and a tracking system produces one thing: a position, on a coordinate model of the pitch, for every player, every official and the ball, repeated many times per second from kick-off to the final whistle. That is the entire product. Distance covered is not measured, it is calculated from those positions. Sprint counts are not measured either. Neither is possession share, defensive line height, or the space behind a full-back. All of it is arithmetic performed after the fact on a list of coordinates, which means every one of those numbers inherits whatever the capture got wrong.
What a tracking system produces before anyone calls it a statistic
The output file is unglamorous. Each row holds a timestamp, an object identifier, and two or three numbers giving that object's location. Twenty-two players plus a ball, sampled twenty-five times a second across ninety minutes of playing time, is around three million rows before anybody has computed anything at all.
Nothing in that file knows what football is. It does not know a pass happened. It does not know which team is attacking, which direction they are attacking in, or that the match has restarted after a corner. It knows that object seventeen was at one set of coordinates at 43:12.40 and a slightly different set at 43:12.44. Every football concept you eventually read on a dashboard has been inferred from changes in those coordinates, or imported from a second feed and stapled on.
That distinction matters more than any single metric. Positional data is close to raw sensing, and its errors are physical: a lens, a lighting condition, a body hidden behind another body. Everything built on top of it is interpretation, and its errors are definitional. Confusing the two is how people end up arguing about camera quality when the disagreement is really about where somebody drew a threshold.
- 12Dedicated tracking cameras installed
- 29Body points tracked per player
- 50Position samples per second
- 500Ball sensor readings per second
Figures published by FIFA for the semi-automated offside system used at the 2022 World Cup. Structural specification, not a performance claim.
Optical camera arrays: how a stadium becomes a coordinate grid
An optical system starts with several synchronised cameras mounted high around the ground, each covering an overlapping slice of the playing surface. Before a ball is kicked, each camera is calibrated against fixed reference points whose real-world positions are known: the corners of the penalty area, the intersections of the touchline and the goal line, the penalty spot. Calibration is what converts a pixel in an image into a location on the pitch, and it is the reason the painted lines are more useful to the system than they are to the players.
With calibration done, each frame goes through three stages. The software separates moving foreground from static background, so twenty-odd human-shaped blobs are picked out of a green field. It then decides which blob is which, assigning a persistent identity to each so that the player at one end of the pitch in minute two is recognisably the same object in minute eighty. Finally it converts each blob's screen position into pitch coordinates using the calibration, and combines the views from multiple cameras so that a player seen by three of them resolves to one location rather than three.
Stage two is where the difficulty lives. Detecting that something is moving is easy. Keeping track of who it is through ninety minutes of players in identical shirts crossing, colliding, falling over each other and disappearing behind a wall of bodies at a corner is genuinely hard. This is the same class of problem that camera systems in other sports have to solve, and the parallels with how tracking cameras work on a basketball court are close, though a football pitch is roughly ten times the area with fewer cameras per square metre.
Two properties of optical capture are worth holding onto. It sees everybody, including the opposition, the substitutes warming up behind the goal, and the officials. And it requires nothing to be attached to a player, which is why it is the only option for capturing an away team's movement in a competitive fixture.
Wearable pods: satellites, local beacons and the sensor that never lies about impact
The other family of capture puts the sensor on the player. A pod sits in a vest between the shoulder blades, and inside it are usually three separate technologies that people lazily collapse into the single word "GPS".
The satellite receiver is the part that earns the name. It fixes a position from global navigation satellites, typically reporting around ten times a second. Its accuracy outdoors is decent for distance and speed over sustained running and considerably worse for short, sharp movement, because a change of direction completed in a third of a second falls between too few samples to be described properly.
The local positioning layer replaces satellites with fixed transmitters around the stadium. Because the beacons are close, the geometry is much better than anything satellites offer, and indoor venues and roofed stadiums become workable. The trade is a permanent installation, which means the system belongs to the venue rather than travelling with the squad.
The inertial layer is the one that adds information neither cameras nor satellites can supply. Accelerometers and gyroscopes sample at rates in the hundreds per second and record the forces going through the player rather than their location. Collisions, jump take-offs and landings, and the deceleration load of stopping hard all show up here and are close to invisible in a positional feed. Under the IFAB's Law 4, wearable devices are permitted in matches provided the competition organiser is satisfied they are not dangerous, and FIFA's quality programme certifies devices against a safety test as well as, since it was extended, a positional accuracy test.
Clubs run both families, and not because they are indecisive. Wearables give the medical and conditioning staff a load record for their own players in training, where no camera array exists. Optical gives the analysts a complete picture of both teams in matches, where wearables cover only half the pitch.
Sampling rate is the setting that quietly decides what you can measure
Frame rate sounds like a technical footnote and behaves like a constraint on the entire analysis.
Consider what happens between two samples. At ten samples a second, one sample lasts a tenth of a second, and a player running at seven metres per second travels seventy centimetres inside it. At twenty-five, the gap is twenty-eight centimetres. At fifty, it is fourteen. Anything shorter than the gap between samples is not slow at that rate, it is invisible. A step-over is invisible. A body feint that shifts a defender's weight for a fifth of a second is invisible. The instant a boot meets a ball is not captured at all, only inferred from the frames either side of it.
That is why the ball sensor in FIFA's offside installation reports five hundred times a second while the limb cameras report fifty. The kick instant is the one quantity in the whole system where a hundredth of a second changes the answer, so it gets its own instrument.
The trade is storage, processing and the amount of noise you have to filter. Higher rates capture faster events and also capture more jitter, so raw coordinates are always smoothed before anyone calculates a speed from them. Smoothing is a choice made by the supplier, it is rarely documented in detail, and it has a direct effect on peak speeds, because an aggressive filter shaves the top off every acceleration.
Centroid tracking and skeletal tracking answer different questions
For most of the history of football tracking, a player was a dot. Centroid tracking collapses a human being into one point, usually the centre of mass projected down onto the pitch surface, and that dot is entirely sufficient for a large class of questions. How far did he run. How fast. How compact was the defensive block. How much space sat between the midfield line and the back four when the ball was played forward.
Skeletal tracking replaces the dot with a set of points on the body: shoulders, elbows, hips, knees, ankles, feet, head. FIFA's published specification for the 2022 offside system tracked twenty-nine such points on each player, chosen to include every limb and extremity that offside can legally turn on.
The extra points buy three things a dot cannot provide. They give body orientation, so the system knows which way a defender was facing when the pass went in behind, which is the difference between a well-timed run and a lucky one. They give posture, so a jump, a slide and a stumble are distinguishable rather than all being a dot that briefly slowed down. And they give extremity positions, which is what an offside judgement actually needs, because the law concerns the parts of the body a player can score with rather than the middle of the torso.
Skeletal capture is also considerably more expensive to compute and more fragile in crowds, since inferring an occluded knee is harder than inferring an occluded person. Most clubs get skeletal output for a subset of situations rather than continuously across the match.
Occlusion, identity swaps and the operator who fixes them
Every optical tracking supplier has a room of people watching the output, and the reason is occlusion.
When two players in the same shirt cross at close range, the system briefly sees one blob where there were two. When they separate, it has to decide which identity goes with which body, and it will sometimes get it wrong. Alone, a single swap looks harmless. In practice, a swapped identity means the wrong player is credited with the next forty seconds of running, the wrong player is recorded as making the recovery run, and the two players' physical totals for the match are both wrong in opposite directions.
Corners and defensive walls are the worst cases: fifteen bodies inside a few square metres, most of them jumping, several of them making deliberate contact. Goalmouth scrambles are close behind. A busy match may need dozens of manual corrections, and a supplier's real quality is measured less by its raw detection rate than by how quickly and reliably those corrections get made before the file is delivered.
This has a consequence people rarely think about. Live tracking, delivered during the match, has had less correction applied than the file a club receives afterwards. If a live graphic and a post-match report disagree, the post-match report is usually the one that has been cleaned.
How event data is joined to positional data
Positional data knows where everybody was. It does not know that a pass was attempted, that a tackle was made, or that a shot was blocked. Those come from event data, which is a separate feed produced by a different process, historically human annotators watching video and increasingly a mixture of automated detection and human confirmation.
An event record is a small set of fields: a timestamp, a type, a player, a pitch location, an outcome, and a handful of qualifiers such as whether a pass was headed or which foot took the shot. Event data is rich in football meaning and blind to everyone not directly involved. It will tell you a pass was played from one point to another. It will not tell you that there were four opponents inside the passing lane and a striker making a decoy run to drag a centre-back away.
Joining the two feeds is what makes both useful, and it is more delicate than it sounds.
- Capture both feeds independentlyThe camera array or pod system produces a continuous stream of coordinates. Separately, annotators and detection software produce a list of discrete football events with their own timestamps. Neither feed can see the other while it is being made.
- Align the two clocksBoth streams carry a match clock, but they rarely agree to the fraction of a second. A coarse alignment is done from obvious shared moments such as kick-off and restarts after a stoppage, which removes whole-second offsets.
- Refine each event against the ballCoarse alignment is not enough for a pass. The join is tightened by searching the positional stream near the stated event time for the instant the ball velocity changed sharply, close to the player credited with the action. That instant, not the annotator timestamp, becomes the event time.
- Attach the full frame to the eventWith the timing fixed, the event inherits the coordinates of all twenty-two players and the ball at that instant. A pass record stops being two points and becomes a complete snapshot of the pitch at the moment it was played.
- Derive the contextual featuresFrom that snapshot, the interesting quantities are computed: how many opponents sat between the ball and the goal, how much distance the nearest defender had to close, which team-mates were available and how well covered each one was.
- Validate and reject the bad joinsSome events will not match anything sensible in the positional stream, usually because the ball was occluded or the event was mis-stamped. Those are flagged rather than forced, because a wrongly joined event is worse than a missing one.
The generic sequence any provider or club analytics group has to run. Naming and ordering vary between suppliers; the steps do not.
Step three is the one that decides whether the whole dataset is trustworthy. Get it wrong by three tenths of a second and every pass is recorded with the defenders in the positions they held before the ball was struck rather than at the moment of striking, which systematically flatters or punishes players depending on how their teams move. A model trained on that data will learn the offset rather than the football.
Physical output metrics: what distance, sprints and accelerations really count
The physical numbers are the ones that escape into broadcast and press coverage, and each one is a definition wearing a measurement's clothes.
Total distance is the sum of the small distances between consecutive smoothed positions. It is the most robust physical figure available, because random errors in the coordinates go in both directions and largely cancel across the tens of thousands of steps in the sum.
High-speed running distance is the distance covered while speed exceeds a threshold. There is no universal threshold. Different suppliers, leagues and sports science groups use different values, and moving the line by half a metre per second can change a player's total for the match by a wide margin.
Sprint count requires a threshold and usually a minimum duration above it, so that a single noisy sample does not register as a sprint. Both parameters are the supplier's choice.
Accelerations and decelerations count the occasions the rate of change of speed crosses a threshold. These are the most sensitive numbers of all, because acceleration is the second derivative of position, and every layer of smoothing applied to the coordinates changes it.
Metabolic or high-intensity load applies a model that converts speed and acceleration into an energy cost. It is a model output, not a measurement, and different models disagree.
The practical rule that follows is short. Compare physical numbers only within one supplier, one definition and one competition. A club's internal season-on-season comparison is meaningful. A comparison of one league's published sprint totals against another's is very often a comparison of two thresholds.
Where physical output numbers stop being trustworthy
Beyond the threshold problem there are three failure modes worth naming, because all three are routinely ignored in public analysis.
The first is that physical output is a consequence of tactics rather than a measure of effort. A side that sits deep and defends its own box will cover less ground than one pressing high, and both may be executing their instructions perfectly. A midfielder in a possession-dominant team runs less than a midfielder chasing the game. Comparing raw distance between players in different systems is comparing job descriptions.
The second is that the numbers say nothing about whether the running was useful. A hundred metres of sprinting into a lane the ball was never going to be played into is recorded identically to a hundred metres of sprinting that pinned two defenders. This is exactly the gap that possession-adjusted and context-aware metrics were invented to close, and it is why pressing intensity measured through opposition passes tells you something different from raw distance covered.
The third is aggregation. A ninety-minute total hides its own distribution. Two players with identical distance figures can have completely different match profiles, one working steadily throughout and one producing repeated maximal efforts inside two ten-minute spells. For injury risk and for substitutions, the distribution is the part that matters, and it is the part the headline number destroys.
Semi-automated offside runs on the same feed, with two things added
The officiating application of tracking is the one the public sees, and it reuses everything already described with two additions.
The first addition is a much denser body model. Offside turns on the position of the parts of the body with which a player can legally play the ball, so a centroid is useless and even a coarse skeleton is not enough. Hence the twenty-nine tracked points in the published FIFA specification, sampled fifty times a second.
The second is a precise kick instant. Offside is judged at the moment the ball is played by a team-mate, and everything else in the calculation is downstream of finding that moment. A sensor inside the ball reporting five hundred times a second, combined with the positional stream, replaces an operator scrubbing through broadcast frames to pick the closest one.
What the system delivers is narrower than the branding implies. It produces limb positions and a kick time, and applies a geometric comparison to them. It does not rule on whether the attacker was interfering with play, whether a defender deliberately played the ball, or whether the phase was over. Those questions belong to the law rather than to the sensor, and the officials still answer them. The mechanics of that handover, and what the video officials do with an automated alert, are set out in the piece on what a video review actually involves, and the underlying legal tests are covered in the offside law and its three involvement clauses.
The honest description is that tracking removed a measurement problem from officiating and left the judgement problem exactly where it was. That is a real improvement, and it is not the same as automation.
What tracking made possible that event data alone could not
Three families of analysis exist only because somebody knows where all twenty-two players were.
Space models. Given every position and velocity, you can compute which team would reach each part of the pitch first, and turn a snapshot into a map of controlled territory. That map is what makes it possible to say a pass was available rather than merely that it was not attempted, and to evaluate the decision rather than only the execution.
Off-ball valuation. A striker who occupies two centre-backs so a midfielder can receive between the lines never appears in an event feed. In a positional feed, the effect of his movement on the defenders around him is directly observable, which is the beginning of paying players for work that produces no touch.
Defensive structure. Line height, block width, the compactness of the two banks, and the moments those distances break are all measurable continuously rather than being described in adjectives. The measurable version of what pressing structures such as gegenpressing are trying to achieve lives here.
Note what is common to all three. None of them is a better version of an existing statistic. They are questions that could not previously be asked at all, which is the more interesting kind of change and also the reason clubs guard their positional feeds more carefully than their event data.
Why two providers hand a club two different numbers for the same match
A club running two systems will get two answers, and the gap is rarely a scandal. It has four ordinary sources.
Capture differs: a camera array and a pod system are measuring different physical quantities and will disagree most on short, sharp movements. Smoothing differs: how much filtering is applied before speed is calculated changes every peak. Definitions differ: thresholds, minimum durations and what counts as playing time are supplier choices. And correction differs: how much manual identity fixing has been applied before delivery varies by supplier and by turnaround time.
FIFA's quality programme exists partly to make the first of those comparable, by testing certified systems for positional and velocity accuracy against a reference. It does not standardise the second, third and fourth, which is why two certified systems can both be accurate and still report different sprint counts. Accuracy is about the coordinates. Sprint counts are about the arithmetic performed afterwards, and no certificate covers that.
Tracking the ball is a separate problem from tracking the players
Everything so far has treated the ball as one more object in the feed. It is not, and the reasons are worth setting out, because ball position is the input almost every derived football metric depends on.
A ball is small, it moves far faster than any player, and it spends a good deal of the match in the air where the flat-pitch assumptions that make player tracking tractable stop holding. A player is always standing on the ground, so a single camera plus a calibrated pitch plane is enough to place him. A ball at head height is somewhere along a line stretching away from the lens, and pinning it down needs at least two cameras that can both see it at the same instant, which is precisely what a crowded penalty box denies you.
It is also the object most likely to be hidden. It is obscured by legs at every tackle, by bodies at every corner, and by the netting, the crossbar and the goalkeeper at the one moment everybody cares most about. Suppliers fill those gaps by interpolating a plausible flight path between the last confident sighting and the next one, which is a reasonable engineering answer and also means some of the ball trajectory in any feed is inferred rather than seen.
That is the argument for putting a sensor inside the ball. An inertial unit at the centre of the ball does not care about lenses or crowds, and it reports the instant of contact directly rather than requiring the software to guess it from a change in direction. The published FIFA specification for the 2022 offside system pairs exactly that sensor, reporting five hundred times a second, with the optical limb tracking, because the two technologies fail in different places and cover each other.
Who is allowed to see the feed, and why the public sees so little
The last structural fact about tracking data is commercial rather than technical, and it explains most of the strange shape of public football analysis.
Event data is licensed widely. It is the reason shot maps, pass maps and expected goals graphics appear on television and in newspapers, and it is why the shot-quality models built on that feed became the first advanced statistic ordinary supporters recognised. Positional data is licensed narrowly. A competition typically contracts a single supplier to install the arrays, and the resulting feed goes to the clubs in that competition, to the league itself, and to a small number of partners.
Two consequences follow. The first is that almost every public tactical argument is conducted with event data alone, which is to say with a record of what happened to the ball and no record of where anybody else was standing. The second is that clubs have a genuine and durable information advantage over the outside world, which is unusual in modern sport and is the reason recruitment departments will pay for positional access before they pay for almost anything else.
There is a third, less obvious consequence. Because the feed is scarce, the methods built on it are developed inside clubs, published rarely, and validated by nobody outside the building. A public metric gets torn apart by strangers within a fortnight of release. A private space-control model gets torn apart by whoever the club employs, and that is the entire quality control process.
Checking a tracking claim before you repeat it
Five questions handle almost everything that goes wrong in public discussion of tracking numbers.
Which feed is this from? Positional, event, or a model built on both. A number described as tracking data is often a model output with a tracking input, and models can be wrong in ways sensors cannot.
What is the threshold? For any sprint, high-intensity or acceleration figure, the definition is the number. Without it the figure cannot be compared with anything.
Was it a full match? Per-ninety normalisation of physical output is not innocent, because a substitute's twenty minutes are not a third of a starter's hour at the same intensity. Rate figures flatter late substitutes.
Does the tactic explain it? Before attributing a low distance figure to effort, check where the team defends and how much of the ball it has. Most surprising physical numbers are system effects.
Live or delivered? A number quoted during a broadcast has had less correction applied than the file the club sees the next morning.
Get into the habit of asking those and the tracking discourse becomes much quieter, because a large share of the claims made about players from physical data do not survive the second question. The rest of the tactics, rules and money explainers sit in the football archive, and the wider collection of sports explainers is in the blog index.
Common questions
What is football tracking data?
It is a stream of positions for all twenty-two players, the officials and the ball, recorded many times per second for the whole match and stored as coordinates on a model of the pitch. On its own it contains no events, no passes and no shots, only where everything was at every sampled instant. Everything else a club reports from it, including distance covered and sprint counts, is calculated afterwards from those coordinates.
Is optical tracking better than GPS pods?
They answer different questions. Optical arrays cover every player on the pitch including the opposition, which is the only way to model shape and space, but they need a fixed installation and struggle when bodies overlap. Wearable pods only cover the players wearing them, which in a match means one squad, but they carry accelerometers that capture collisions and short accelerations a camera resolves poorly.
What frame rate does football tracking run at?
Optical match tracking has commonly run at around twenty-five samples per second, which matches broadcast video, and dedicated officiating installations run faster. FIFA published a specification for the 2022 semi-automated offside system using fifty samples per second for limb positions and a ball sensor reporting five hundred times per second. Higher rates matter most for events that happen inside a fraction of a second, such as the exact instant a ball is kicked.
What is the difference between centroid and skeletal tracking?
Centroid tracking reduces each player to a single point, usually the centre of mass projected onto the pitch, which is enough for distances, speeds and team shape. Skeletal tracking follows many separate points on the body, so it knows where an arm, a knee or a shoulder was rather than only where the player was. Offside decisions and body-orientation analysis need the second kind; distance covered does not.
How is event data joined to tracking data?
Both feeds carry a match clock, so the join starts by aligning the two clocks, then refines each event by looking for the moment in the positional stream when the ball motion changed near the player credited with the action. Once aligned, every event inherits the position of all twenty-two players at that instant. The alignment is the fragile step, and a systematic offset of a few tenths of a second corrupts everything computed downstream.
Are distance and sprint numbers reliable?
Total distance is the most robust physical figure because errors in both directions cancel over ninety minutes. Sprint counts, high-intensity distance and acceleration counts are far less comparable, because each depends on a speed or acceleration threshold and a minimum duration that the supplier chooses. Two systems watching the same match can report very different sprint totals without either being wrong, which is why the threshold matters more than the number.
Filed under Football·football · data · analytics · technology · officiating · sports science