AI PoweredConnect
Enter your Axis camera’s IP address or the video URL of a compatible IP camera. Keep a live view of your timing point.
THE NEXT GENERATION OF TRACKPIXEL
TrackPixel V2
AI PoweredKeep the image.
Find the passing.
Video backup designed for race timekeepers. From the camera stream to the competitor’s photo, TrackPixel V2 brings together tools to retrieve, read and verify a passing.
ETERNYTIME / TRACKPIXEL V2YOUR TIMING POINT, IN PICTURES
A PASSING AND ITS VISUAL RECORD
One interface.
All your passings.
Capture & read
Set a detection area and sensitivity. Motion triggers a photo; recognition attempts to read visible numbers. Uncertain readings remain available for review.
Verify & export
Find photos in the list, confirm numbers and export passings to Excel. Return to live video without losing your work.
A photo provides a visual reference. Timestamp accuracy depends on the camera, stream and synchronisation; the photo alone is not certified race timing.
SHARED ETERNYTIME LEARNING
A community of timekeepers.
Experience shared across races.
Timing just one or two events a year? Your observations still matter. An angled quad number, a distinctive motorcycle font or a bib in difficult lighting provides a different example. By pooling readable, validated crops, users build a dataset that better reflects real racing conditions.
Every contribution matters
You do not need thousands of photos on your own. A few relevant, correctly labelled crops can add variety. Quality matters more than quantity.
Diversity strengthens the dataset
Disciplines, number styles, viewing angles and lighting bring challenges that individual timekeepers may not encounter at their own events.
Shared benefits, measured
The aim is to share selected improvements with participating installations. Adding an example does not instantly improve every reading: a new AI model must be trained, compared and validated before publication.
From a number on the track to shared experience
You stay in control
Confirm the number and its crop in one window. The correction enters local visual memory immediately, without Internet or a Train button. A number region with identical pixels can then be recalled. Moved, blurred or different plates still rely on OCR; memory does not guarantee recognition at the next passage.
Correct once. Share automatically.
Save your community connection once and allow sharing. Authorised crops are then sent automatically without a second approval. Offline crops stay on the PC and uploads resume when the network returns, retrying about once per minute while TrackPixel is open. Full photos and camera credentials are not sent.
Train, then compare
With the camera disconnected and settings windows closed, TrackPixel automatically tries training when at least 30 examples are available and the dataset has changed. Diversity and evaluation requirements still apply. An insufficient model never replaces the active AI. Signed, evaluated community models are checked and activated automatically outside capture; a successful check is repeated after six hours.
Controlled model sharing
The server collects crops. Community computation runs on an open Eternytime Windows installation with an administrator key and its camera disconnected. It imports contributions, trains, evaluates and publishes only eligible improvements automatically. Web hosting alone does not train models: data waits if that installation is off. A new shared model is not guaranteed after every correction.
Racing continues offline.
Recognition runs locally on the PC. Sharing examples and obtaining models require Internet access and can take place between events. Every proposed number still needs confirmation.
Correct once. Share automatically.
Save your community connection once and allow sharing. Authorised crops are then sent automatically without a second approval. Offline crops stay on the PC and uploads resume when the network returns, retrying about once per minute while TrackPixel is open. Full photos and camera credentials are not sent.
With the camera disconnected and settings windows closed, TrackPixel automatically tries training when at least 30 examples are available and the dataset has changed. Diversity and evaluation requirements still apply. An insufficient model never replaces the active AI. Signed, evaluated community models are checked and activated automatically outside capture; a successful check is repeated after six hours. The server collects crops. Community computation runs on an open Eternytime Windows installation with an administrator key and its camera disconnected. It imports contributions, trains, evaluates and publishes only eligible improvements automatically. Web hosting alone does not train models: data waits if that installation is off. A new shared model is not guaranteed after every correction.
Taking part in practice
What is sent?
The validated number crop, its digits, discipline and a hash of the session name to group examples. This channel sends neither the full photo nor camera credentials. Crop tightly, without faces or helmets.
Does it learn by itself during a race?
Confirm the number and its crop in one window. The correction enters local visual memory immediately, without Internet or a Train button. A number region with identical pixels can then be recalled. Moved, blurred or different plates still rely on OCR; memory does not guarantee recognition at the next passage. With the camera disconnected and settings windows closed, TrackPixel automatically tries training when at least 30 examples are available and the dataset has changed. Diversity and evaluation requirements still apply. An insufficient model never replaces the active AI. Signed, evaluated community models are checked and activated automatically outside capture; a successful check is repeated after six hours. The server collects crops. Community computation runs on an open Eternytime Windows installation with an administrator key and its camera disconnected. It imports contributions, trains, evaluates and publishes only eligible improvements automatically. Web hosting alone does not train models: data waits if that installation is off. A new shared model is not guaranteed after every correction.
Do I need constant Internet access or to contribute at every event?
Save your community connection once and allow sharing. Authorised crops are then sent automatically without a second approval. Offline crops stay on the PC and uploads resume when the network returns, retrying about once per minute while TrackPixel is open. Full photos and camera credentials are not sent.
Can every number be read reliably?
No. Hidden, blurred or muddy numbers may remain unreadable. The dataset grows from real contributions and improvements need measurement. Proposed numbers still require verification.
THE REALITIES OF RACING
Motorcycles. Quads. Karting.
Cycling. Cars.
Angled numbers, mud and changing light: validated examples need to reflect actual race conditions.
Mass starts are a challenge of their own.
In a dense pack, counting vehicles and identifying competitors are separate tasks. Helmets, colours and silhouettes can help suggest a match at the next passing.
A visual resemblance does not confirm a number. Matches need verification; a fully hidden vehicle may require a second camera view.
LET’S DISCUSS YOUR SETUP
Your camera.
Your race environment.
Explore how TrackPixel V2 can fit into your video setup and race timing operation.