[Camera Support]: LPR on Frigate+

[haroldboom]

Describe the problem you are having

Hi all, I had LPR working well on the non-Frigate+ config however, since I switched to native LPR support for Yolov9 on Frigate+ I cant get the plate text to be recognized. I enabled debugging and got this in my log output however, no debug images were saved in the media folder.

I have cut my config down to just the camera that does LPR and my LPR config.

Version

.17 beta 2

What browser(s) are you using?

No response

Frigate config file

Camera Config:
  Driveway:
    ffmpeg:
      output_args:
        record: preset-record-generic-audio-aac
      inputs:
        - path: rtsp://127.0.0.1:8554/Driveway
          roles:
            - record
        - path: rtsp://127.0.0.1:8554/Driveway_sub
          roles:
            - detect
    detect:
      width: 1280
      height: 720
      fps: 6
    objects:
      track:
        - person
        - car
        - motorcycle
        - license_plate
      filters:
        person:
          threshold: 0.85
          min_area: 500
        car:
          threshold: 0.8
          mask: 
            1280,0,1280,720,788,720,789,410,742,205,683,158,556,130,399,145,412,179,297,251,186,280,0,344,0,0
        licence_plate:
          threshold: 0.55
    snapshots:
      enabled: true
      bounding_box: true
      required_zones:
        - driveway_person
        - driveway_car
      retain:
        default: 30
    zones:
      driveway_person:
        coordinates: 
          1280,720,1280,166,873,62,691,71,553,91,386,112,21,178,73,387,213,720
        objects:
          - person
      driveway_car:
        coordinates: 
          1280,720,900,720,866,462,780,238,669,149,606,140,549,130,402,146,420,168,369,207,220,276,0,348,0,720
        objects:
          - car
          - motorcycle
          - license_plate
    ui:
      order: 2
    review:
      genai:
        enabled: true
      alerts:
        required_zones:
          - driveway_person
          - driveway_car

LPR config:

lpr:
  enabled: true
  recognition_threshold: 0.5
  debug_save_plates: true

Relevant Frigate log output

debug | 2026-01-14 09:08:17 | frigate.data_processing.common.license_plate.mixin | Driveway: Running plate recognition for id: 1768342087.358719-3u6wg5.
debug | 2026-01-14 09:08:17 | frigate.data_processing.common.license_plate.mixin | Driveway: No boxes found by OCR detector model
debug | 2026-01-14 09:08:17 | frigate.data_processing.common.license_plate.mixin | Driveway: No text detected
debug | 2026-01-14 09:08:17 | frigate.data_processing.common.license_plate.mixin | Driveway: No attributes to parse.
info  | 2026-01-14 09:08:17 | logging                                            | Last message repeated 2 times
debug | 2026-01-14 09:08:21 | frigate.data_processing.common.license_plate.mixin | Driveway: Running plate recognition for id: 1768342087.358719-3u6wg5.
debug | 2026-01-14 09:08:22 | frigate.data_processing.common.license_plate.mixin | Driveway: No boxes found by OCR detector model
debug | 2026-01-14 09:08:22 | frigate.data_processing.common.license_plate.mixin | Driveway: No text detected
debug | 2026-01-14 09:08:22 | frigate.data_processing.common.license_plate.mixin | Driveway: No attributes to parse.
info  | 2026-01-14 09:08:22 | logging                                            | Last message repeated 1 times

Relevant go2rtc log output

N/A

FFprobe output from your camera

N/A

Frigate stats

No response

Operating system

UNRAID

Install method

Docker Compose

Object Detector

OpenVino

Network connection

Wired

Camera make and model

Hikvision

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

Let me analyze your configuration. I notice a typo in your config that's likely causing the issue:

In your camera's object filters, you have:

licence_plate:
  threshold: 0.55

This should be license_plate (with an 's' instead of 'c')(1). The incorrect spelling means your filter isn't being applied to the license_plate object type.

When using a Frigate+ model that natively detects license_plate, you need to ensure license_plate is added to your list of objects to track (which you have correctly done) and that any filters use the correct spelling(1)(2).

Additionally, a few observations about your setup:

1. Your recognition_threshold is set to 0.5, which is lower than the default of 0.9(1). While this makes recognition less strict, it won't help if the OCR detector isn't finding text boxes in the first place.

2. The log shows "No boxes found by OCR detector model," which suggests the text detection step is failing before character recognition even runs(1).

3. You might want to try adjusting the min_area setting for license plates if they're appearing small in your frame. The default is 1000 pixels (area measurement)(1).

4. Consider testing the enhancement parameter at the camera level, starting around 5 if plates are visible but characters aren't being recognized(1). However, based on your logs, the issue appears to be at the text detection stage rather than character recognition.

Fix the typo first and see if that resolves your issue. If plates still aren't being recognized after correcting the spelling, you may need to adjust min_area or review whether the plates are large/clear enough in your detect stream.

📚 Sources:

1. License Plate Recognition (LPR)
2. Models

---

Have another question? Just tag @inkeep.

[hawkeye217]

Based on your screenshot above, it looks like the blurriness/pixelation of your plate could be preventing the LPR subsystem from reading it correctly. Nothing looks incorrect with your config and the LPR subsystem seems to be working. You could try increasing the bitrate of the stream in your camera firmware. 1280x720 should be fine for a resolution, but if you can't increase bitrate, you could try increasing resolution.

[haroldboom]

Based on your screenshot above, it looks like the blurriness/pixelation of your plate could be preventing the LPR subsystem from reading it correctly. Nothing looks incorrect with your config and the LPR subsystem seems to be working. You could try increasing the bitrate of the stream in your camera firmware. 1280x720 should be fine for a resolution, but if you can't increase bitrate, you could try increasing resolution.

Thanks for your help mate, I can increase the bitrate. Current bitrate on the substream is 2048, any suggestion on what I should increase it to?

[hawkeye217]

2048 should be plenty. I don't think increasing it would change your issue, but you could try doubling it.

One thing you'll want to have a look at is how the car is being detected. With a Frigate+ model, license_plate is an attribute of car, so if the car stops being detected, the plate will also no longer be detected (and thus not be run through LPR).

One suggestion for debugging would be to create a "dummy camera" and run an exported clip through it with the same settings as your driveway camera (and a manual setting of the detect resolution to match). This will give you a real-time view of exactly what Frigate is seeing and can help you change any settings as needed.

https://docs-dev.frigate.video/troubleshooting/dummy-camera

[haroldboom]

2048 should be plenty. I don't think increasing it would change your issue, but you could try doubling it.

One thing you'll want to have a look at is how the car is being detected. With a Frigate+ model, license_plate is an attribute of car, so if the car stops being detected, the plate will also no longer be detected (and thus not be run through LPR).

One suggestion for debugging would be to create a "dummy camera" and run an exported clip through it with the same settings as your driveway camera (and a manual setting of the detect resolution to match). This will give you a real-time view of exactly what Frigate is seeing and can help you change any settings as needed.

https://docs-dev.frigate.video/troubleshooting/dummy-camera

I have doubled the bitrate and I also increased the minimum size.

I have a hunch but is it possible that the plate is recognized while the vehicle is moving and it is trying to process a blurry image that's far away and then doesn't run when the vehicle stops? I have this in my log

2026-01-14 09:08:07.512133986 [2026-01-14 09:08:07] frigate.data_processing.common.license_plate.mixin DEBUG : Driveway: Skipping LPR for non-stationary car object 1768342087.358719-3u6wg5 with no position changes. (Detected in 4 concurrent frames, threshold to run is 5 frames)

I will play around with the setting a bit more and then if that fails I will try the dummy camera. Thanks again

[hawkeye217]

That log message means that the car has not been marked as stationary internally yet, but it has not had any position changes. This is the initial state for a new tracked object, so it's generally not unexpected to see that in the logs. But if your car was detected, lost, and then re-detected, that log may indicate that kind of situation.

[codex70]

I've also just switched from the free version to Frigate+ and have had to switch back to the standard lpr functionality. I'm also running 17.2 beta.

Frigate+ is not picking up license plates that are easily read with the normal version. If I remove/comment out -license_plate it works well, as soon as I add it back in again, it doesn't pick up the plates.

My relevant settings are:

objects:
  track:
    - person
    - bicycle
    - car
    - motorcycle
    - bird
    - cat
    - dog
    - face
    - license_plate
  filters:
    cat:
      threshold: 0.5
      min_score: 0.4
    person:
      threshold: 0.8
    bird:
      threshold: 0.8
      min_score: 0.6
    license_plate:
      threshold: 0.5
      min_score: 0.4

lpr:
  enabled: true
  model_size: small
  recognition_threshold: 0.5
  match_distance: 2
  enhancement: 0
  debug_save_plates: true

  known_plates:
    My Car:
      - XXXXXXXX
      
logger:
  default: info
  logs:
    frigate.data_processing.common.license_plate: debug

[hawkeye217]

I downloaded the clip just fine, thanks. And yes, that inference time is quite problematic. What CPU are you running?

[haroldboom]

I downloaded the clip just fine, thanks. And yes, that inference time is quite problematic. What CPU are you running?

AMD Ryzen 9 9950X and my GPU is a B580 Battlemage. I am running Frigate in a docker on Unraid. Should have tons of power, CPU is at ~10% usage most of the time

[hawkeye217]

I just ran your clip through the base model with the detect resolution set at 1920x1080 and it looks like the LPR code is working just fine. So it looks to be something going on with your particular setup. Let's see if we can get to the bottom of the high inference time.

[hawkeye217]

lpr:
  enabled: true
  model_size: small
  device: CPU
  recognition_threshold: 0.5
  enhancement: 0
  debug_save_plates: true

Try this as your LPR config and see if you get different results.

[haroldboom]

lpr:
  enabled: true
  model_size: small
  device: CPU
  recognition_threshold: 0.5
  enhancement: 0
  debug_save_plates: true

Try this as your LPR config and see if you get different results.

Done, I will report back on how it goes. Thanks!

[haroldboom]

lpr:
  enabled: true
  model_size: small
  device: CPU
  recognition_threshold: 0.5
  enhancement: 0
  debug_save_plates: true

Try this as your LPR config and see if you get different results.

It seems to be working, one successful vehicle plate identified so far. That’s strange though right because aren’t the settings you gave me default besides the threshold and debug images?

Update: working perfectly for me thanks for your help mate I will mark as solved.

[hawkeye217]

Great. The device key actually defaults to None, so setting it as CPU was the fix for you here. Another contributor changed the default a few months ago, so the docs need to be updated.

Glad it's working for you now, happy to help!

[codex70]

Just to give you some feedback, I went back through the hundreds of images I'd uploaded and made sure they all hade license plates added for vehicles. It looks like that has done the trick, there haven't been many cars at the gate to test, but it's definitely picked up my car since, so looks like it's working as designed. Thanks again for the help