Skip to main content

Frigate as a detection source

Crumb's own detectors work on raw pixel motion; they don't know the difference between a person, a car, and a blowing branch. If you already run a dedicated object detector on your own hardware, Crumb can use it in two ways.

This is entirely optional and entirely bring-your-own: Crumb does not bundle or run an object detector itself, and pixel motion detection keeps working the same whether or not an integration is configured. See Integrations for the setup steps and what data flows where.

As timeline enrichment (alongside pixel motion)​

By default, an integrated detector's events (person, car, package, and so on, whatever labels it produces) show up as their own icons on the same timeline as pixel motion, distinguishable at a glance. Pixel motion still drives recording; the detections are added context. This is additive and changes nothing about how recording is triggered.

As the recording trigger (per camera)​

You can also let object detections drive recording for a specific camera, in place of pixel motion. Each camera has a motion source that is either pixel (the default, Crumb's own analysis) or frigate (recording is triggered by the detector's object events). Crumb translates an object appearing and leaving into the same start and stop signal the pixel pipeline emits, so recording, pre-roll, and post-roll behave identically; only the trigger differs.

The reason to do this is precision. On a camera pointed at a busy street or moving foliage, pixel motion fires constantly on things you don't care about, whereas object detection only fires on an actual person or vehicle. Switching just that camera to the frigate source removes the nuisance recordings.

The trade-off, read this before switching​

Object-triggered recording is precise, not fail-safe. If the detector does not report an object, Crumb records nothing for that camera, and its timeline shows nothing for that window:

  • Pixel motion is fail-safe: any motion records, so it over-records but rarely misses.
  • Object triggering records only what the detector classifies, so it can miss real events the detector does not catch: low light, an object type it does not track, a detection below its confidence threshold, heavy occlusion, or the detector being down.

For a security camera, missing a real event is usually worse than an extra recording. So treat this as a per-camera decision. Use the frigate source on nuisance-prone cameras where precision matters, and keep pixel motion, or continuous recording, on cameras where you cannot afford to miss anything. If you want both guarantees at once, run the camera on continuous recording and use the object detections purely for timeline highlights: you never miss footage, and you still get a smart, searchable timeline.