Frigate: Open-source network video recorder with real-time AI object detection

Original Article ↗ Hacker News Discussion ↗

Frigate, an open-source network video recorder with real-time AI object detection, is attracting attention as a powerful alternative to commercial NVR systems like Blue Iris, especially for users running Linux or low-power hardware. Commenters report strong results with person and vehicle detection, flexible integrations (Home Assistant, MQTT, go2rtc, OpenVINO, Coral TPUs, Rockchip NPUs), and both event-based and continuous recording, though some still find its playback UI and mobile-friendly UX less polished than mature proprietary tools. Much of the debate centers on hardware choices (PoE vs Wi‑Fi cameras, GPU/TPU accelerators), the reliability of motion and object detection models, and long-term viability of accelerators like Google Coral.

Overall Sentiment & Use Cases

  • Many commenters are enthusiastic; several call Frigate the best open-source NVR they’ve used.
  • Typical setups: home security, remote dwellings, door monitoring, license plate capture, wildlife/bear alerts, and fun automations via Home Assistant.
  • Some run Frigate alongside a traditional NVR for redundancy and continuous archival footage.

Frigate+ and Custom Models

  • Users of Frigate+ report noticeably better accuracy, especially at night, and far fewer false positives.
  • Skepticism exists about the added value since Frigate already supports user-supplied custom models for free.
  • Some want to see more concrete comparisons and understand how much Frigate+ improves over well-tuned YOLO models.

Continuous vs Event-Based Recording

  • Frigate supports 24/7 recording with per‑camera and per‑object retention policies; some users are satisfied.
  • Others say the continuous playback UI is still weak (scrubbing, variable speed, searching), so they stick with or supplement with tools like Blue Iris or Moonfire NVR.
  • Several argue continuous recording plus event tags is industry best practice, since motion/object detection can miss events.

Motion & Object Detection Quality

  • Earlier OpenCV-based motion detection is criticized as noisy; version 0.13 introduces a rewritten algorithm that testers say is much better.
  • Public object-detection models (often ImageNet-based) are seen as ill-suited to typical ceiling/long-shot security views, leading to misses.
  • Some report near-perfect detection in their setups; others stress that unnoticed false negatives are hard to measure.
  • Custom YOLOv8 and specialized people-detection models are reported to work very well in some deployments.

Hardware, Acceleration & Cameras

  • Coral TPUs help but are hard to source and poorly supported on some Linux distros; building libraries can be painful.
  • Newer Frigate versions support Intel OpenVINO on iGPUs and some Rockchip NPUs, reducing reliance on Coral.
  • Raspberry Pi 4, small Intel boxes, M1 Macs, and GPUs (e.g., Tesla P4) are all reported viable, depending on camera count.
  • Strong preference for wired PoE cameras over Wi‑Fi due to bandwidth, reliability, and jamming risks; others use Wi‑Fi successfully when wiring isn’t feasible.
  • go2rtc integration improves Reolink reliability, allows protocol conversion, and supports two-way audio.

Integrations, UX & Legal Concerns

  • Deep integration with Home Assistant, MQTT, and other tools (Scrypted, HomeKit) is a major selling point; notifications and automations are highly customizable.
  • A polished, family-friendly mobile app is seen as missing compared with Blue Iris.
  • Some raise privacy/legal concerns about residential video surveillance, which vary by jurisdiction and are not fully resolved in the thread.