The Second Edition of the Analytics Cup is now openSign Up
LATEST NEWS
All News & Analysis

The New Benchmark for XY Tracking

Single-camera tracking data that raises the bar for quality and consistency.

We are excited to launch a significant upgrade to our single-camera XY Tracking data. Built on our 10 years of experience, this update sets the new industry benchmark, offering numerous quality and feature upgrades that both enrich existing workflows and provide the basis for even more advanced analysis.

Watch the video below for an overview or read on for the full detail.

The three key enhancements can be summarised as greater accuracy in player and ball tracking, more granular quality filtering and new data points and delivery methods.

Greater Accuracy

The headline improvement in this new tracking version is the increased precision of the outputs at both a player and ball level across every key measure we track, including a significant upgrade in the accuracy of player speed.

Improved Accuracy on Every Measure We Track
Versus previous Tracking XY version (V3, 2024)
Comparison to ground-truth optical tracking data
Player position
accuracy
(within ±2 m)
97.0%
Up from 95.5%, on visible players.
Player speed
accuracy
(within ±1 km/h)
89.1%
Up from 74.3% at frame level, on visible players.
Ball position
accuracy
(within ±2 m)
94.7%
Up from 90.9%, on visible ball.
Ball in play
agreement
98.3%
Up from 96.0%: more than half of the wrong calls fixed.

This improvement has been driven by three key Computer Vision upgrades:

  • Enhanced visual descriptors: Improved differentiation of players and team, meaning more players are identified and associated with the correct identity.
  • Improved ball tracking: An updated model that provides more accurate and realistic ball trajectories, particularly on high balls and in relation to player touches.
  • Better homography: An improved model, leading to better 3D real-world coordinates estimation from the 2D image.


And the uplift is clear across both visible and non-visible players and ball.

Visible Players and Ball

The percentage improvements noted above are mirrored in a marked improvement in the mean absolute error (MAE) on both player and ball position, and player speeds.

Validation Study: Visible players and ball
Based on an analysis of 1,137 matches
V3 (2024)
V4 (2026)
MAE on player positions
1.03 m
0.71 m
Accuracy on player positions(< 2 m to GT)
95.5%
97.0%
MAE on player speeds
0.85 km/h
0.52 km/h
Accuracy on player speeds(< 1 km/h to GT)
74.3%
89.1%
MAE on player acceleration
0.35 m/s²
0.21 m/s²
Accuracy on player acceleration(< 0.5 m/s² to GT)
79.1%
91.3%
MAE on ball positions
1.03 m
0.65 m

Non-Visible Players and Ball

The estimation of position and movement for players outside the view of the camera is one of the key challenges facing any single-camera tracking provider.

We were already confident that our extrapolation model was one of the most accurate in the industry, but we have taken advantage of the quality improvements elsewhere in this new version to produce a full rework of our model.

The updated model leverages Computer Vision tracking outputs on visible objects (ball, player, main referee) to predict likely movements by bridging the non-visible frame gaps for every player and outputting more consistent trajectories.

The result is increased accuracy across all key measures.

Validation Study: Non-visible players and ball
Based on an analysis of 1,137 matches
V3 (2024)
V4 (2026)
MAE on player positions
3.40 m
2.55 m
Accuracy on player positions(< 4 m to GT)
75.1%
82.8%
MAE on player speeds
1.92 km/h
1.54 km/h
Accuracy on player speeds(< 2 km/h to GT)
65.6%
74.0%
MAE on player acceleration
0.47 m/s²
0.35 m/s²
Accuracy on player acceleration(< 1 m/s² to GT)
89.1%
94.0%
MAE on ball positions
4.58 m
3.61 m

Use Case: Connecting Tracking to Event Data

One of the key use cases that our customers highlight is the ability to combine tracking data with standard event data for enhanced contextual analysis and modelling.

This matching of data types is usually achieved on a frame and positional basis, and fails if player or ball position strays beyond a certain threshold.

The increased accuracy of our upgraded tracking data makes this process more reliable, with significant gains in precision across important event types, including passes, shots and corners.

Validation Study: Matching event data to tracking frames
Based on an analysis of 1,137 matches and 1.33M events
V3 (2024)
V4 (2026)
Accuracy on pass events(< 2 m within 0.2s)
85.9%
94.9%
Accuracy on shot events(< 2 m within 0.2s)
69.6%
87.1%
Accuracy on corner events(< 2 m within 0.2s)
40.3%
87.9%

To put the improvement in concrete terms: almost nine out of every 10 shots can now be matched to event data, up from seven out of 10 in our previous tracking version.

More Granular Quality Filtering

Alongside the quality upgrades, we are also providing more granular control of the tracking data you choose to ingest and analyse.

We have built new frame-level error predictions for the position and speed of players and the ball, allowing you to filter out the data of specific players or passages of play depending on your use case and quality thresholds.

Previously, quality information was only provided at a match level. That often led to matches with difficult tracking conditions being discarded entirely. Now, you have the information you need to filter out or alter your approach for low-quality data, whilst keeping the usable data from those matches for your normal analysis.

That might mean:

  • Filtering out the first five minutes of a match where flare smoke lingered in the air after kickoff.
  • Discounting players on the far side of the pitch from footage filmed at a low camera angle. 
  • Removing sequences in which sunlight directly into the lens makes player and ball recognition difficult.
    ‍

The flexibility is now there to keep and discard as you choose.

This frame-level detail has also powered an overhaul of our existing match-level player and ball quality indexes to provide a clearer representation of the overall tracking quality of a match.

New Data Points and Delivery Methods

For the first time, we are serving frame-level speed and acceleration values as part of our tracking data, allowing you to immediately derive actionable information without having to compute those values yourself.

The latest version also provides two more quality of life improvements that have been requested by various customers to more easily integrate our tracking data into their pipelines. The tracking data can now be delivered at 25 FPS, in addition to the existing 10 FPS feed, and we have added a new delivery format: Parquet.

We also now provide tracking data for the main referee in each match, opening up additional use cases for leagues, federations and refereeing organisations.

Quality & Consistency

All of these improvements are available across every one of the 170+ competitions that we cover worldwide, providing a consistent platform for scouting and benchmarking.

Tracking data is the base on which all of our solutions are built, so these advancements not only provide a higher-quality product to our XY Tracking customers, but they will also feed into the models that power our Physical Data and Game Intelligence solutions.

When the accuracy of our underlying tracking data increases, all SkillCorner customers benefit.

Get Started

All of these quality and feature upgrades are available from Thursday 19th November to existing SkillCorner XY Tracking customers via API and our Web Query Tool.

Interested in adding tracking data to your own data stack? Get in touch today.

Articoli Correlati

Sblocca il vero valore dei dati di tracking

richiedere una demo