TL;DR: Cloud home cameras keep making headlines for exposed footage. Self-hosting with Frigate NVR fixes that, but it has its own trap: it can look like it works while reading nothing. I measured it. At a 640x360 detection size, Frigate tracked 16 vehicles and read 0 plates. At 1280x720, same footage, it read 6 plates, 3 of them useful. Here is what I learned, what I tested, and what I did not.
Your footage is only as safe as someone else's server
Cloud camera vendors have a recurring problem: footage ending up where it should not.
- Wyze (Feb 2024): After an outage, about 13,000 users were shown camera thumbnails and events belonging to other people (Engadget).
- Bitsight (June 2025): researchers found roughly 40,000 internet-exposed cameras streaming to anyone with a browser (The Register, Bitsight).
- Meari (May 2026): flaws in Meari's cloud platform left over a million baby monitors and security cameras exposed (eSecurity Planet).
- Ring: settled with the US FTC over claims that employees and contractors had improper access to customers' videos (Malwarebytes).
The pattern is the point: when video leaves your house, you inherit someone else's security posture.
Frigate is an MIT-licensed NVR that runs object detection locally on your IP cameras. Nothing has to leave your network.
The trap: local does not mean it works
Plate reading (LPR) in Frigate is a chain of four steps: detect a vehicle, find the plate, read the characters, match against what you know. Any link can fail, and failure is silent. You get a dashboard with no errors and no plates.
The Frigate docs say it plainly: if you cannot read the plate yourself in the image, Frigate will not either.
What I measured
All results are from a laptop (Intel Core i9-12900H, Iris Xe), using looped video footage, not a live camera.
| Test | Setup | Result |
|---|---|---|
| T1 (Frigate 0.17.2) | Two wide street clips, 640x360 | 253 cars tracked, 0 real plates. Only "reads" were ADVANC and ADVANCE painted on a bus |
| T3 (Frigate 0.18.2) | Daylight 1080p road footage with readable Nigerian plates, 112 s looped 3 times, detect at 640x360 | 16 vehicles, 0 plate reads |
| T3 (same footage) | Detect at 1280x720 | 15 vehicles, 6 plate reads, 3 useful |
Of the 6 reads at 1280x720: the same plate read correctly twice, once more with an 8 misread as a B, plus a 4-character junk read, the camera's on-screen "SDCard" text, and a likely misread. Recognition took about 118 ms per plate, above the 100 ms the docs warn about.
The lesson: LPR runs on the detect stream, so detect size decides whether plates can be read at all.
Three settings that clean up the noise
The junk reads above are fixable. These come from the Frigate LPR docs (the format regex below is the docs example; use your country's pattern):
lpr:
enabled: True
min_plate_length: 4
format: "^[A-Z]{2}[0-9]{2} [A-Z]{3}$"
match_distance: 1
known_plates:
Delivery Van:
- "ABC1234"
-
min_plate_lengthandformatfilter out overlay text and junk reads. -
match_distance: 1lets a B/8 misread still match a known plate.
I did not run this exact config against the T3 footage in this post; the claim that it would remove the junk is my reasoning from the settings' documented behaviour.
Hardware, per the docs
- Avoid Wi-Fi cameras; streams drop and disconnect.
- Use H.264 video with AAC audio for the widest compatibility.
- LPR needs a CPU with AVX and AVX2 and at least 4 GB of RAM.
- The docs name the Beelink EQ13 (Intel N100) as a favourite for several 1080p cameras with low to medium activity. Coral is no longer recommended for new installs.
- Turn on snapshots or recording for the plate camera. In my tests with both off, Frigate tracked vehicles but saved none.
Local still needs patching
Frigate 0.18 added a lot, and 0.18.2 carried security fixes. Self-hosting moves the responsibility to you: keep it updated, isolate cameras on their own network, and use a VPN for remote access rather than opening ports.
What I tested, and what I did not
I label every claim in the book:
- TESTED: run by me on the stated hardware and version.
- DOCUMENTED: from the official docs, not yet tested by me.
- ILLUSTRATIVE / UNTESTED / FIELD NOTE: examples, unverified items, and lessons from installs.
My numbers come from a laptop and video clips. Your camera, lighting and scene will differ. I have not tested a live outdoor camera at night for this post.
The book
I turned this into a practical weekend handbook for Frigate 0.18 on a small Intel mini PC, no GPU or Coral needed: Home Cameras You Can Trust. There is a free sample on the page.
Frigate, Frigate NVR and Frigate+ are trademarks of Frigate, Inc. This guide and post are independent and not affiliated with, sponsored by or endorsed by Frigate, Inc.
Discussion
- Have you caught an LPR setup "working" while reading nothing? What was the cause?
- Would you trade cloud convenience for local control on cameras covering your front door?
- What detect resolution do you run, and what does it cost your CPU?
Sources
- Frigate documentation (hardware, license plate recognition): https://docs.frigate.video
- Incident reports are linked inline above (Engadget, The Register, Bitsight, eSecurity Planet, Malwarebytes).
Top comments (0)