Skip to content

License Plate Recognition (ANPR) — LicensePlateRecognizerBlock

LicensePlateRecognizerBlock reads vehicle license plates with a specialized two-stage pipeline: a dedicated license-plate detector (YOLO) locates plates in the frame, then a plate-specific OCR model reads the characters of each cropped plate. Both models come from the FastALPR family (MIT): a YOLOv9-T plate detector and a fast-plate-ocr recognition head. A global head covers the USA and 90+ countries; a European head is tuned for EU plates. The recognition model's geometry and alphabet are read from the model itself, so selecting a region is simply a matter of pointing RecognitionModelPath at that head.

graph LR;
    Source-->LicensePlateRecognizerBlock;
    LicensePlateRecognizerBlock-->VideoRendererBlock;
    LicensePlateRecognizerBlock-. OnPlateRecognized .->App[Your app];

Usage

using VisioForge.Core.MediaBlocks.AI;
using VisioForge.Core.Types.X.AI;

var settings = new LicensePlateRecognizerSettings(detectorModelPath, recognitionModelPath)
{
    Provider = OnnxExecutionProvider.Auto,
    DetectionConfidenceThreshold = 0.35f,
    OcrConfidenceThreshold = 0.3f,
    DrawResults = true,
};

var anpr = new LicensePlateRecognizerBlock(settings);
anpr.OnPlateRecognized += (sender, e) =>
{
    foreach (var plate in e.Plates)
    {
        Console.WriteLine($"Plate: {plate.Text} ({plate.Confidence:P0}) at {plate.BoundingBox}");
    }
};

pipeline.Connect(source.Output, anpr.Input);
pipeline.Connect(anpr.Output, videoRenderer.Input);

await pipeline.StartAsync();

Each LicensePlateResult carries the recognized Text (normalized: uppercased, non-alphanumeric characters stripped), the mean Confidence (0..1), the axis-aligned BoundingBox, and the detection Polygon (four OcrPoint vertices), all in source-frame pixels.

Key settings

LicensePlateRecognizerSettings(detectorModelPath, recognitionModelPath):

Property Default Description
DetectorModelPath — License-plate detection ONNX model (FastALPR YOLOv9-T end-to-end). Required.
RecognitionModelPath — Plate-OCR ONNX model (FastALPR fast-plate-ocr head, global or EU). Required.
Provider / DeviceId Auto / 0 ONNX execution provider and hardware device index.
FramesToSkip 0 Skip frames between recognition runs on live video.
DetectionInputSize 640 Square input size for the detection model (dynamic-shape models).
DetectionConfidenceThreshold 0.35 Minimum detector score a plate box must reach.
OcrConfidenceThreshold 0.3 Minimum mean OCR confidence a recognized plate must reach to be reported.
MaxDetections 10 Maximum plates detected per frame.
DrawResults true Draw plate boxes and text into the video frame.
BoxColor / BoxThickness Yellow / 3 Overlay styling.
LabelFontSize 0 0 auto-scales to frame height.

Models and licensing

Both the detector and the recognition head are MIT-licensed FastALPR models; the SDK does not ship model weights in the NuGet package. For higher accuracy on busy scenes with many small or distant plates, run a dedicated general-purpose vehicle detector (for example YOLOObjectDetectorBlock) upstream to crop vehicle regions before ANPR.

Use with VideoCaptureCoreX and MediaPlayerCoreX

var anpr = new LicensePlateRecognizerBlock(settings);
anpr.OnPlateRecognized += Anpr_OnPlateRecognized;

core.Video_Processing_AddBlock(anpr); // before StartAsync (VideoCaptureCoreX)
// player.Video_Processing_AddBlock(anpr); // before OpenAsync/PlayAsync (MediaPlayerCoreX)

await core.StartAsync();

See Using AI blocks with VideoCaptureCoreX and MediaPlayerCoreX for the full processing-block API, insertion order, and lifecycle rules shared by every video AI block.

Use cases

  • Parking access and payment — recognize a plate at a gate camera to open a barrier or start a parking session.
  • Toll and access-control logging — record which plates passed a fixed camera and when.
  • Fleet and yard management — track vehicles entering/leaving a private lot or depot.
  • Traffic enforcement support tooling — flag plates for manual review (final enforcement decisions should always have a human-review step).

Troubleshooting

Symptom Likely cause Fix
Plates aren't detected at all DetectionConfidenceThreshold too high, or plate too small relative to DetectionInputSize Lower DetectionConfidenceThreshold; raise DetectionInputSize for distant/small plates, or crop closer to the vehicle upstream.
Plate detected but text is wrong/empty OcrConfidenceThreshold too high, or wrong regional recognition head Lower OcrConfidenceThreshold; confirm RecognitionModelPath matches your region (global vs. EU head).
Only some plates in a busy scene are reported MaxDetections reached Raise MaxDetections if you expect more than 10 plates per frame.
Plate text includes stray characters Read directly from LicensePlateResult.Text, expecting raw OCR Text is already normalized (uppercased, non-alphanumeric characters stripped) — if you still see noise, check that the correct regional recognition head is loaded.

Frequently Asked Questions

Does this ANPR SDK work outside the USA and Europe?

The FastALPR global recognition head covers the USA and 90+ countries; a separate European head is tuned for EU plates. Point RecognitionModelPath at whichever head matches your target region.

Do I need to train my own plate-detection model?

No — LicensePlateRecognizerBlock uses the FastALPR YOLOv9-T detector and fast-plate-ocr recognition head out of the box; you only need to supply the two .onnx files.

Can I use LicensePlateRecognizerBlock on a wide traffic scene with many vehicles?

Yes, up to MaxDetections plates per frame (default 10, configurable). For very busy scenes with small/distant plates, consider running a vehicle detector (YOLOObjectDetectorBlock) upstream to crop vehicle regions first.

Is license plate data considered personal/biometric data?

License plates are commonly treated as personal data under privacy regulations (though not biometric in the same sense as a face embedding). Review the applicable regulations (GDPR, state ANPR laws, and similar) for your jurisdiction and use case before deploying.

Demos