The AIs are amazing at analyzing images but they can make mistakes. Sometimes they see a
person briefly on the floor to pick up something, which is real but probably not worth an alert. And sometimes
when using night vision, the AIs might mistake a pile of pillows on the floor for a
person.
So we developed Alert Strictness: We ask the AIs to look multiple times, or to
compare with each other, before sending an alert.
Sometimes the best course of action is to move a camera for a better angle, or to
reduce clutter in the monitored area.
System Accuracy
As of October, 2026, we have a growing “corpus” of 204 images that span 3 test
homes. These images show us on the floor in many different places and types of lighting. The
corpus also includes images where we know that no one is on the floor… either the room is
empty, or there’s a large dog, etc.
Here’s how our system performs at spotting us on the floor (called “recall”),
as well as how often it thinks someone is on the floor when no one actually is (“false
positive”):
- At Alert Strictness level 0: Recall is 96.8%; false positives are 3.8%
- At Alert Strictness level 4: Recall is 94.4%; false positives are 0%
Your results may be different. These figures are from our 3 test homes, with our
cameras, rooms, and lighting.