Created: 8/2/2026
Updated: 8/18/2026
fastALPR runs Automatic License Plate Recognition (ALPR) locally, in-process, using the
fast-alpr Python package. Detection and OCR both run
as ONNX models on your own hardware (CPU or GPU) - no external server, API key, or internet
access is required at inference time, unlike components such as CodeProject.AI or Deepstack.
fastALPR is a post processor and works together with an object detector component (e.g.
darknet, yolo, edgetpu). It runs on the bounding boxes of detected vehicles (e.g. car,
truck) and looks for a license plate within them.
Configuration
Configuration example
/config/config.yaml
fastalpr:
license_plate_recognition:
cameras:
viseron_CHANGEME_camera:
labels:
- car
- truck
known_plates:
- ABC123
save_plates: true
min_confidence: 0.6
detector_model: yolo-v9-t-384-license-plate-end2end
ocr_model: cct-xs-v2-global-model
device: auto
License plate recognition domain config.
Camera-specific configuration. All subordinate keys corresponds to the camera_identifier of a configured camera.
Camera identifier. Valid characters are lowercase a-z, numbers and underscores.
A list of labels that when detected will be sent to the post processor. Applies only to this specific camera.
A mask is used to exclude certain areas in the image from post processing.
List of X and Y coordinates to form a polygon
X-coordinate (horizontal axis).
Y-coordinate (vertical axis).
A list of labels that when detected will be sent to the post processor. Applies to all cameras defined under cameras.
List of known license plates. Each plate will have its own sensor.
If set to true, detected license plates will be stored in the database, as well as having a snapshot saved.
Minimum confidence for a license plate detection.
Lowest value: 0
Highest value: 1
Time in seconds before a plate recognition expires.
Which license plate detector model to use. See
models for more information on this.
Valid values:
yolo-v9-s-608-license-plate-end2endyolo-v9-t-640-license-plate-end2endyolo-v9-t-512-license-plate-end2endyolo-v9-t-416-license-plate-end2endyolo-v9-t-384-license-plate-end2endyolo-v9-t-256-license-plate-end2end
Minimum confidence for the detector to consider something a license plate. This is separate from min_confidence, which instead filters the result of the OCR reading.
Lowest value: 0
Highest value: 1
Which OCR model to use to read the characters of a detected license plate. See
models for more information on this.
Valid values:
cct-s-v2-global-modelcct-xs-v2-global-modelcct-s-v1-global-modelcct-xs-v1-global-modelcct-s-relu-v1-global-modelcct-xs-relu-v1-global-modelargentinian-plates-cnn-modelargentinian-plates-cnn-synth-modeleuropean-plates-mobile-vit-v2-modelglobal-plates-mobile-vit-v2-model
Device used to run inference on.
auto lets ONNX Runtime pick the best available provider. cuda requires an NVIDIA GPU with CUDA/cuDNN available, and is only bundled in the amd64-cuda image. On all other images the CUDA provider is unavailable and inference falls back to the CPU.
The OCR model can detect a license plate-shaped region but fail to read any characters from it, returning an empty plate text. This can happen even when min_confidence is set high, since the model's confidence in this case does not reflect whether any characters were actually read. If true, these empty results are discarded instead of being reported as a detection.
License plate recognition
License plate recognition runs as a post processor when a specific object is detected.
You can track known license plates by adding them to the known_plates list in the configuration.
Known plates will be reported as binary sensors.
There is also a sensor entity that reports all the plates that have been detected.
This can be used to trigger actions in other platforms, eg Home Assistant, when an unknown plate is detected.
Models
fastALPR uses two separate ONNX models: a detector model that locates license plates in the
image, and an OCR model that reads the characters of a detected plate.
detector_model is a yolo-v9-*-license-plate-end2end model. The number in the name is the
input resolution the model runs at, e.g. yolo-v9-t-384 resizes its input to 384x384. Higher
resolutions (e.g. yolo-v9-s-608) are more accurate at finding small/distant plates but slower,
while lower resolutions (e.g. yolo-v9-t-256) are faster but may miss smaller plates.
ocr_model reads the characters from the plate crop found by the detector. cct-*-global-model
and global-plates-mobile-vit-v2-model are trained on plates from many regions/formats.
european-plates-mobile-vit-v2-model and argentinian-plates-* are trained on region-specific
plate formats and can be more accurate if your plates match one of those regions. xs and relu
variants trade some accuracy for faster inference.
Troubleshooting
To enable debug logging for
fastalpr, add the following to your
config.yaml/config/config.yaml
logger:
logs:
viseron.components.fastalpr: debug
