The camera that keeps watch for three months on one small battery

What happened

A camera is the hungriest sensor you can put on a battery. Even a small, modern camera chip eats about 13.5 milliwatts when it runs all the time. That sounds tiny. On a small battery it means the device is dead in hours. So today’s battery cameras cheat. They sleep, wake up now and then, take a picture, and sleep again. What do they miss while asleep? Everything that moved.

A team at ETH Zurich in Switzerland asked a different question. What if the camera did not have to be awake to notice motion? They built a node with two eyes. One eye is an “event camera”. It does not take pictures. It only reports when something in its view changes, and it sips power while it waits. The other eye is a normal color camera. It stays off until the first eye says “something moved”.

The test

The node is 50 by 25 by 15 millimeters, about the size of a matchbox. Waiting for motion, it draws 222 microwatts. That is less than a thousandth of the power of a small LED bulb. When the event camera sees motion, it wakes a small processor and the color camera. The processor runs a tiny detection model, with one million settings, that can name 80 kinds of objects. Then it sends the answer over LoRa, a long-range radio that uses very little power.

The whole job, from waking up to sending the answer, takes about half a second and 28.7 millijoules of energy. That is about as much energy as a phone uses to light its screen for a fraction of a second.

The result

At a typical rate of activity, with something moving in front of it about one percent of the day, the node runs for three months on a 1.85 watt-hour battery. A phone battery holds roughly ten times that. On a hard standard test of object detection, the small model scores about 32 percent, and a larger version reaches nearly 40 percent, which is respectable for a chip this small.

Why is it so efficient? This is because the node spends energy in proportion to what happens. A quiet field costs almost nothing. A busy gate costs more, and only while it is busy. Think of a dog asleep with one ear up. Watching costs nothing. Looking costs a lot. Therefore, look only when the ear twitches.

What it means

Most of the world has no mains power where the interesting things happen: the farm gate, the borehole, the telecom tower, the elephant corridor, the warehouse yard. A camera you can place and forget for a season changes what can be watched. The pieces used here, the event camera and the small AI chip, are already on sale. PLACE IT AND FORGET IT.

Business ideas from this paper

1. 1. Season-long cameras for farms, boreholes and towers

   Build a rugged version of this node, sell it with a solar top-up, and let it report “person”, “vehicle” or “animal” over LoRa to a phone. Theft of batteries and fuel at telecom towers is a daily problem across East and West Africa.

   Buyer Farm cooperatives, tower operators, water projects, conservancies. First test Pre-sell ten units to one cooperative or one tower company before you build the housing.

2. 2. A watch-and-alert service, not a device

   Most buyers do not want hardware. They want a message that says “someone is at the gate”. Rent the nodes, run the gateway, and charge per site per month.

   Buyer Security firms and property managers. First test One paid pilot at five sites for 90 days, at a price that covers the hardware in six months.

3. 3. One-industry detection models

   The general model in the paper names 80 everyday objects. A farm needs “goat”, “stray dog” and “person at night”. Train small models for one industry at a time and license them.

   Buyer Device makers and integrators in agriculture, wildlife and logistics. First test Offer three integrators a free model in exchange for field data and a testimonial.

How sure can you be?

This is a preprint, so it has not yet been reviewed by other scientists. The lab is one of the strongest in the world for low-power electronics, and every number comes from measurements on a real device, not a simulation. The battery life is a calculation from those measurements at an assumed one percent activity rate. A busy scene will drain it faster. The detection accuracy is modest. It is enough to say “a person is there”, not to read a number plate.

Source: An Energy-Proportional Multimodal and Context-Aware Vision IoT Node, arXiv preprint, August 2026. arXiv:2608.23192 · arxiv.org (preprint · not yet peer reviewed).

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.