DOVBETA
All projects

Security · Computer vision

Plate Gate

License plate recognition that turns ordinary IP cameras into a smart parking log — every pass on video, every car accounted for.

Status
Live
Year
2026
What we delivered
Computer vision pipeline · Web portal · Video storage · On-premises deployment · Monitoring & backups
clean event per pass instead of dozens of frame reads
1
clean event per pass instead of dozens of frame reads
GPUs or proprietary hardware needed
0
GPUs or proprietary hardware needed
cameras in production, entry and exit
3
cameras in production, entry and exit

The challenge

Residential complexes, business centres and private car parks need to know which cars come and go — but off-the-shelf ANPR systems are expensive, closed and locked to one vendor's hardware. Most sites already have IP cameras on the gates; they just record video nobody watches.

And before any system is trusted to open a barrier, its accuracy has to be proven on the site's own cameras, angles and lighting — headlights at night, glare, infrared switching — not on a vendor's demo footage.

The solution

Plate Gate connects to existing RTSP cameras and runs fully on-premises on open-source models. A YOLOv9 detector finds plates, a tracker follows each car across frames, and an OCR model reads every frame; voting across the whole track turns dozens of noisy reads into a single, confident event per pass, validated against Ukrainian plate formats.

The web portal shows a live journal with a photo and a short video clip of every pass, who is parked right now and for how long, anomalies worth checking, a registry of owners and their cars imported from Excel, and the health of the server and every camera. Operators can verify reads, and a built-in evaluation tool measures real accuracy by day, night, entry and exit.

Key capabilities

  • 01

    Works with existing cameras

    Any RTSP IP camera or DVR channel. No vendor lock-in, no cloud, all data stays on site.

  • 02

    Video of every pass

    A 5–20 s clip starting 2 s before the plate appears, streamed to the browser with seeking and cleaned up automatically.

  • 03

    Who's inside right now

    Parking presence derived from entries and exits, with missed exits and one-character mismatches flagged instead of guessed.

  • 04

    Recognition zones

    Draw the driveway on a live frame per camera; billboards and signs stop producing false reads, no restart needed.

  • 05

    Owner registry

    Owners, units and cars imported from Excel with a dry run; every pass shows who the car belongs to.

  • 06

    Measured accuracy

    Operator verification and an evaluation report by day/night and entry/exit — numbers, not promises.

Screens

Parking: who is inside right now, for how long, and anomalies to review (demo data)
Parking: who is inside right now, for how long, and anomalies to review (demo data)

The result

Plate Gate runs 24/7 in production with three cameras — one entry and two exits — on an ordinary office computer without a GPU.

The architecture already separates "what was seen" from "who is allowed in", so automatic barrier control is a configuration step on top of proven accuracy, not a rewrite.

Technology

  • Python
  • OpenCV
  • ONNX Runtime
  • YOLOv9
  • FastAPI
  • PostgreSQL
  • MinIO
  • go2rtc
  • FFmpeg
  • Docker Compose

Want something like this for your business?

Tell us about your idea or the process you want to automate. We'll propose a solution, a timeline and the first working version.

Next case

Spot Watch

Shows which parking spots in a residential courtyard are free right now — using the IP cameras already on site.