A camera that looks away from people
Nobody wants a camera pointed at them
Most wildlife cameras record everything that moves and leave the rest to you. On a balcony, much of what moves is people. That is fine for a security camera. It is not fine for something meant to watch your tomatoes.
Filtering people on the device
Every photo is checked by a small AI model on the camera itself. A photo of a person is never stored or shown; only animals reach your gallery.
We did not build that model from scratch. Recognising people is one of the most studied problems in small-device AI, and a model trained on millions of images beats anything trained on one balcony. Our work was fitting it into a camera that runs on a small battery, and deciding how cautious it should be.
The hard part: not missing the squirrel
A filter that is too eager starts mistaking animals for people. A squirrel sitting upright, a cat on the railing, a pigeon close to the lens: drop those, and the camera goes quiet exactly when it should act.
So we did not guess. The camera first ran in a silent mode on our testers’ balconies, judging every photo without dropping any, until we had thousands of real, labelled photos. Then we chose the setting that kept every animal in that set. People get filtered; squirrels still get noticed.
Never scaring a person
The worst thing a deterrent can do is play a predator call at someone. So the camera is deliberately reluctant: it needs two photos that agree before it reacts, and a single photo of a person or your own pet is enough to keep it quiet. Without Wi-Fi it adds a check that does not rely on the camera at all: if a known phone or a pet’s Bluetooth tag is nearby, it stays silent.