Monitor your security cameras with locally processed AI

Original Article ↗ Hacker News Discussion ↗

Overall impressions & use cases

  • Many commenters run Frigate for years and consider it the best local object-detecting NVR they’ve used, especially compared to Ring, Eufy, Tapo cloud systems.
  • Common uses: door/driveway alerts, package delivery, theft deterrence, monitoring workshops/garages, livestock and wildlife watching, monitoring ponds and rural properties, and checking on vulnerable family members or pets.
  • Some enjoy the “hidden world” of nighttime wildlife (foxes, raccoons, deer, insects), others focus on reducing nuisance alerts (“no more small animals waking me up”).

Accuracy, models, and Frigate+

  • Default model is generally praised but not perfect: false positives on toys, flags, garden clutter, trees, and odd shapes; occasional misses at night or with headlights.
  • Paid Frigate+ subscription is seen by supporters as fair value: funds training, keeps improved models indefinitely, and greatly reduces false positives for them.
  • Others dislike sending sensitive footage for cloud training or feel locked out of easy custom-model workflows; one person says the default model “sucks” for their use and they won’t upload medical-related footage.

Hardware, accelerators, and codecs

  • Coral TPU: widely used and still works well with Frigate, but many complain it’s effectively “abandonware” (ancient Python deps, EOL distros). Some say modern CPUs/iGPUs rival its performance.
  • Alternatives: Intel iGPU/OpenVINO, Arc GPUs, Nvidia, Hailo, Rockchip NPUs, Orange Pi 5; several report low CPU usage even with multiple cameras.
  • Some emphasize that accelerators aren’t necessary for small camera counts if using low-res MJPEG substreams and OpenVINO.
  • Codecs: debate over H.265. Some report smooth end-to-end H.265; others on Linux/Firefox can’t view clips without expensive transcoding and strongly recommend H.264 for that stack.

Integrations, workflows & other software

  • Tight Home Assistant integration (often via MQTT) is a major draw; common patterns include sending image alerts to Telegram/Pushover and using HA automations for alarms.
  • Some use Frigate standalone; others combine with Doubletake + Compreface for face recognition. Frigate has recently added built-in face recognition.
  • Alternatives mentioned: Camect (local but closed box), Scrypted, Ubiquiti Protect with AI Key/Port, ZoneMinder, Motion/MotionEye plus YOLO, go2rtc and MediaMTX, OpenIPC firmware, UniFi NVR hardware.
  • A few find Frigate “overweight” or configuration-heavy, noting config format churn and difficulty when trying to detect objects not in the built-in classes.

Privacy, networking & camera choices

  • Strong preference for local-only systems: many isolate cheap IP cams (Tapo, Eufy, Reolink, AliExpress brands) on VLANs, static IPs, and firewall rules, blocking all internet access.
  • There is skepticism toward cloud vendors (Ring, Eufy): complaints of ads inside apps, cloud clip failures, past security issues, and police or employee access.
  • Some warn that Wi‑Fi-capable cameras could in theory connect to nearby open networks, but most consider VLAN isolation sufficient in practice.

Adversarial bypass & limitations

  • Thread jokes about bypassing detectors with signs, costumes, boxes (“Metal Gear” style), or adversarial patterns; others note Frigate’s two-stage pipeline (motion via OpenCV + object detection) limits such attacks.
  • Serious remark that adversarial examples and IR flashlights can fool vision systems; motion masks are often needed to avoid tree/fence false positives.

Psychological impact & “why cameras?” debate

  • One line of discussion argues that home cameras increase anxiety and perceived insecurity, especially in safe suburbs.
  • Many push back, citing deterrence, evidence for disputes (neighbors, pets, deliveries), monitoring remote/vacant properties, and non-security uses (wildlife, plants, checking if colleagues are in, storm damage).
  • Several emphasise using automation so you don’t “monitor” feeds, you just get rare, meaningful alerts—reducing rather than increasing cognitive load.

Usability, cost & terminology

  • DIY “stacks” often use PoE 4K AliExpress cameras, Pi or mini-PC (N100, NUC, old i5/i7) plus a small SSD/HDD; Coral or Intel iGPU if needed.
  • Some would prefer a turnkey consumer box with Frigate-like capabilities; lack of a simple appliance is seen as a barrier for time-constrained users.
  • Long, heated sub-thread complains that the Frigate site uses “NVR” without expansion; others argue it’s standard in CCTV, but critics see it as needless jargon and gatekeeping.