Recording only people, dogs or cats

Settings › Motion › What starts a recording has two choices:

• Any kind of movement (the default): a person, a pet, a curtain or a shadow can all start a recording. Alerts still say when a person, dog or cat was seen.
• Only people, dogs or cats, then tick which: motion records only when one of those is in the frame. Tick only Cats and a person walking past records nothing.

The sentence under the choice says exactly what will and won't record, and the main window's Motion panel shows the same thing. Untick the last one and it goes back to Any kind of movement.

This choice is only about what moved. How much movement counts is the sensitivity, and where it counts is the areas you draw; both apply either way.

Choosing Only… can discard real events: anything it doesn't recognise is ignored. A clip you delete costs a moment; one that was never recorded can't be recovered.

Dogs and cats. Apple's on-device recogniser knows exactly these two animals, and tells them apart. It needs the animal reasonably visible: one that is small in a wide shot, mostly hidden, or in dark infrared footage will often be missed. It can't tell breeds apart; to have alerts name your own pets, add them under Pets. A pet lying still makes no motion, so nothing is checked until it moves. To record on barking or meowing, choose Only these sounds under Settings › Sound and tick them.

People. What follows is how the person check decides.

It looks for a human upper body — head, shoulders and torso. It is not a face detector, so it does not need to see a face. Concretely:

• Facing the camera, in side profile, or the back of someone's head — all detected. Facial features aren't required.
• Head and shoulders only, as when someone leans into frame — detected.
• Waist up at an angle — detected.
• Legs or feet with no torso visible — not detected. There is no upper body to find.
• Someone mostly hidden behind furniture, or very small in a wide shot — unreliable.
• Photographs, posters and reflections of people — often detected, because they look like people.

The upper-body model is used deliberately rather than the full-body one, which needs to see legs and therefore fails for anyone behind a desk, a sofa or a kitchen counter — that is, most indoor framing.

The failure mode worth understanding is asymmetric: the filter will occasionally discard a real event, and it never invents one. If you would rather over-record than miss something, leave it off. If a check fails outright, the event is kept rather than dropped.

Cost: it runs only while motion is happening, at most twice a second, rather than continuously. On an Intel Mac there's no Neural Engine, so it's noticeably more expensive there — which is why it defaults off on slower Intel machines.