We've just released version 4.0 of Driblab's Vision Tracking System. This update brings meaningful improvements across three core areas: the richness of per-frame data, the accuracy of jersey number recognition, and the reliability of team detection. Each of these areas has been reworked with a clear goal in mind: delivering tracking data that is more complete, more accurate, and more useful for the analysts and organisations that rely on it. Here's what's new.
Richer tracking data available per frame
The tracking output has been extended with new kinematic features for every player, including speed and acceleration in both X and Y directions, giving analysts a precise, directional understanding of movement across the pitch. This level of detail makes it possible to go beyond positional snapshots and start capturing the physical dynamics behind every run, press, and repositioning.
Ball height estimation has been added, enabling a more complete 3D representation of match events and bringing tracking data closer to reflecting the full physical reality of the game. Whether it's a long diagonal pass, a cross into the box, or a through ball played along the ground, the system can now account for the vertical dimension of the ball's trajectory.
Ball velocity and acceleration are now captured as dedicated metrics, adding a new layer of context to every moment the ball is in play. We'll soon be publishing a dedicated article exploring this feature in depth.
These additions increase the analytical depth of every frame and open the door to more advanced performance and tactical analyses, enabling more sophisticated models built on a richer and more complete data foundation.
.png)
Jersey number recognition improvements
Accurate player identification is fundamental to the value of tracking data, and version 4.0 raises the bar here considerably. A new recognition model has been trained on a significantly larger and more diverse dataset, improving performance across the wide range of broadcast conditions found in real match footage — from top-tier productions to lower-budget feeds with varying image quality.
The model is now more robust in challenging scenarios, including low-resolution broadcasts, motion blur, and difficult camera angles. Precisely the situations where previous versions were most likely to produce errors or gaps in player identification.
.gif)
An additional post-processing and validation stage has been introduced, filtering out unrealistic predictions and reducing false positives before results are finalised. This extra layer of quality control ensures that the data reaching analysts has already been screened for the most common sources of noise.
The combined improvements deliver a 4.5% increase in jersey number recognition accuracy, a gain that translates directly into cleaner, more reliable player-level data across every match processed.
Enhanced team detection through improved colour extraction
Knowing which player belongs to which team may sound like a solved problem, but kit similarity and variable broadcast lighting make it a persistent challenge in any tracking system. Version 4.0 addresses this with a fully redesigned colour extraction module built to obtain more reliable visual signatures from player kits, even in conditions where colour perception is distorted by shadows, floodlights, or compression artefacts.
The analysis now goes beyond the jersey, incorporating information from the shorts to help distinguish teams whose shirt colours are very similar. This seemingly simple addition makes a significant difference in matches where the standard jersey-based approach struggles to separate the two sides reliably.
The improved pipeline delivers a 10.5% increase in team colour identification accuracy, leading to more reliable team assignment throughout the match and reducing the kind of misattribution errors that can propagate through downstream analyses.
The improved pipeline delivers a 10.5% increase in team colour identification accuracy, leading to more reliable team assignment throughout the match and reducing the kind of misattribution errors that can propagate through downstream analyses.
Version 4.0 reflects Driblab's ongoing commitment to pushing the boundaries of what tracking data can capture and how reliably it can do so. Each of these improvements is designed not just as a technical upgrade, but as a direct contribution to the quality of the insights that clubs, analysts, and researchers can extract from match footage. Better inputs mean better models, better reports, and ultimately better decisions on and off the pitch.
These updates are already available across Driblab's tracking pipeline.
If you'd like to know more about how version 4.0 can support your work, get in touch with our team.
ICYMI: This is how our Tracking works!
Learn everything about our Tracking technology.
.png)
