CaissonBlog

Insula is in preview, a camera system that follows what it sees

Insula, also called camOS, pans, tilts and zooms to keep a person framed, flags faces it has never seen, reads number plates and raises an event when someone steps into an armed zone. It is free.

Published
3 min
To read
Insula
Project

Insula is the camera system Caisson has been building into a pan, tilt and zoom camera. It is also called camOS. It is in preview now, and like every Caisson product it is free.

The idea is that a camera should do something with what it sees. Most home cameras record a fixed view and send a notification when pixels change. Insula moves the lens to keep a person in frame, works out whether it knows them, reads the plate on a car that pulls up, and only raises an event when someone is somewhere they should not be.

The preview film. Every shot of the yard is Insula's own footage, with the tracking boxes it drew live.

Following

Insula can sweep the yard on its own until it finds someone, then lock on. If it loses them for 30 seconds it goes back to sweeping. Tap a person on the live picture and the camera pans, tilts and zooms to keep them framed; tap anywhere else and it centres that point. Drawing a box on the picture sends the lens there, and pinching works too.

Framing follows two rules in People mode. A face takes 22% of the frame height, and a person who is moving keeps at least 1 m of margin around them, which leaves room for the next step before the camera moves again.

Auto-track modes
Off, People, Vehicles or Plates. Multi follows people or vehicles, whichever is there, and sends the lens hunting for the plate of any vehicle it follows.
Presets
Save the current view to a slot, then tap its name to go back there.
Ignored regions
Detections whose centre falls inside an ignored region are dropped before tracking. A vehicles-only region still lets people in front of it through.

Recognition

Faces are compared against a gallery, using the top three scores. On real footage from the camera, the best impostor scored 0.33 and the median genuine match scored 0.72. That leaves a wide gap for the match threshold to sit in.

A face it has never seen is flagged as soon as it appears and gets its own group, built up photo by photo, so one stranger who walks past three times is one entry and not three. Train a face to give it a name. When the camera recognises someone who lives there looking at it, it answers with that person's own small gesture: a nod, a wiggle or a circle.

Insula reads number plates, parked cars included. A vehicle whose plate it cannot read is described by colour and likely model instead.

Zones and alerts

Walk a zone and the camera follows your footsteps to draw it, or click its corners on the picture. Each zone can be armed on a schedule. A person whose feet are inside an armed zone counts as occupied: Insula records an event and sets a HomeKit sensor, so the rest of the house can react to it, and a desktop notification arrives with a live snapshot.

Privacy mode

With privacy mode on, Insula keeps no pictures of events at all. Names, plates, times and recognition still work, because face fingerprints are computed in memory and never written out as images. With privacy off, snapshots are kept for a number of days you choose, the records stay after the pictures are gone, and stored event images can be deleted in one step.

Getting the preview

Insula is in preview and free. To try it, or to report a fault, email hello@caissonai.co. More detail is on the Insula project page.

Written by Caisson AI. Questions or corrections go to hello@caissonai.co.