An OBS Studio filter that automatically detects and blurs license plates in your video feed in real time. Designed for IRL / driving streamers who want an additional layer of privacy on top of manually hiding plates.

Not a guarantee. This plugin is an additional layer of protection, not a sole safeguard. Computer vision models miss things — especially plates that are small, heavily angled, motion-blurred, or partially occluded. Always test your setup, keep using your existing privacy habits, and don't rely on this plugin to save you from a mistake.

• Add the License Plate Blur filter to any video source (webcam, game capture, screen capture, ...).
• The filter introduces a configurable delay (default 3 s, adjustable 1–5 s). This is intentional: it gives the AI plenty of time to run on every single frame at full 1080p with no skipping. 3 s is imperceptible to streaming viewers and normal for livestreaming.
• A YOLOv9-t model from fast-alpr runs on every buffered frame and outputs plate bounding boxes.
• A GPU shader Gaussian-blurs each detected region (with a safety padding) before the delayed frame is emitted.

The plugin never reads plate text. It only locates plates and blurs them.

Binary installers will be published on the GitHub releases page once v0.1.0 CI has produced them. Until then, build from source (below).

Pre-built packages, when available:

• Windows — .exe installer
• macOS — .pkg (universal)
• Linux (Ubuntu) — .deb

Pre-built installers bundle the YOLO model, so there is no separate model download step after install.

Requires CMake 3.28+, a C++20 compiler, and the OBS plugin deps (pulled automatically by the preset). Two extra things are needed that the preset does not fetch for you:

1. A prebuilt ONNX Runtime matching the pin in .github/workflows/plugin.yml (currently v1.23.2). Download the platform-appropriate archive, extract it, and point -DONNXRUNTIME_ROOT=... (or $ENV:ONNXRUNTIME_ROOT) at the extracted directory. When unset, CMake looks at .deps_vendor/onnxruntime by default.

2. The YOLOv9-t ONNX weights, which are fetched by:

   # macOS / Linux
   scripts/fetch_model.sh
   
   # Windows (PowerShell)
   powershell -ExecutionPolicy Bypass -File scripts/fetch_model.ps1

   Run this once from the repo root before cmake --preset ...; the weights are picked up at configure time and copied into the plugin bundle.

Then:

# macOS (universal)
cmake --preset macos
cmake --build --preset macos

# Windows (Developer PowerShell)
cmake --preset windows-x64
cmake --build --preset windows-x64 --config RelWithDebInfo

# Linux (libobs-dev / libopencv-dev / libcurl-dev from apt)
cmake --preset ubuntu-x86_64
cmake --build --preset ubuntu-x86_64

• Discrete GPU: ideal. Inference on a 384×384 input with YOLOv9-t is <5 ms on most modern GPUs, so 60 fps source with 3 s buffer uses ~1.3 GB VRAM for the frame ring.
• Integrated GPU: works, but consider shortening the delay to 1–2 s and using CPU inference with 2–4 threads.
• CPU-only: possible at 30 fps; expect 15–40 ms per frame depending on silicon.

• Uses the generic fast-alpr model. Accuracy on US/CA plates is good but not perfect; international plates are hit-or-miss.
• Audio is not auto-delayed (add a matching OBS audio delay filter manually).
• No per-frame tracking/smoothing — each frame's detections are independent. This means occasional one-frame detection gaps may briefly unblur a plate. (Lower the confidence threshold to compensate.)
• Memory usage scales with delay × resolution × fps. A 5 s delay at 1080p60 holds ~2.5 GB of frame data in RAM.

This plugin is a fork of royshil/obs-backgroundremoval with its portrait-segmentation logic replaced by license plate detection. The original plugin's cross-platform CMake, CI, and ONNX Runtime wiring are reused directly. Thanks to Roy Shilkrot and Kaito Udagawa for that scaffolding.

The plate detector is fast-alpr's pretrained YOLOv9-t end2end ONNX model.

SPDX-FileCopyrightText: 2021-2026 Roy Shilkrot roy.shil@gmail.com
SPDX-FileCopyrightText: 2023-2026 Kaito Udagawa umireon@kaito.tokyo
SPDX-FileCopyrightText: 2026 Felipe Reyes

SPDX-License-Identifier: GPL-3.0-or-later