Presence sensing in installations is a solved problem with bad solutions. PIR sensors are cheap and crude — they tell you something moved, not where or how many. Cameras with computer vision are precise and bring a privacy conversation you may not want to have. Depth cameras work and need a computer.
Long-wave infrared sits in an interesting gap, and a new compact camera design makes it more accessible: pairing the STM32N6 microcontroller with Lynred’s ATI320 thermal sensor.
The pairing
LWIR — long-wave infrared, roughly 8–14µm — is thermal imaging proper. Not night vision, which amplifies visible and near-infrared light and needs some illumination. LWIR detects emitted heat, so it works in total darkness with no illuminator at all.
The STM32N6 is the notable half. It’s ST’s microcontroller line with an integrated neural processing unit — meaning inference runs on the MCU rather than on a host computer. Pairing an NPU-equipped microcontroller with a thermal sensor gives you a self-contained device that can both see heat and interpret what it’s seeing, without a Pi or laptop in the loop.
ATI320 indicates a 320-pixel-class sensor array. Low resolution by camera standards, entirely adequate for detecting and tracking warm bodies in a space.
Why thermal is underused in interactive work
Four properties that no other cheap sensor combines:
It works in complete darkness. No IR illuminator, no visible light. For pieces that are meant to be dark — most projection and light work — this removes the usual conflict between sensing and aesthetics. A depth camera’s IR floodlight is itself a visible artefact in a dark room.
It doesn’t identify anyone. At 320-class resolution, a thermal image of a person is a warm blob with a shape. You can count people, locate them, track movement and infer posture. You cannot recognise a face. For installations in public or institutional contexts, a sensor that structurally cannot capture identity is a far easier conversation with a venue, a commissioner or a legal team than a camera plus a promise about what you do with the footage.
It segments people from the background trivially. Bodies are warmer than rooms. The separation that computer vision achieves with a trained model comes almost free thermally — a threshold gets you most of the way. That’s why an MCU-class device is sufficient.
It sees through some materials. Thin fabric, some plastics, darkness, smoke. Useful for sensing through a piece’s own surface rather than mounting a visible camera.
The honest limits
Thermal is not a general-purpose replacement, and the failure cases are specific.
Ambient temperature matters enormously. On a hot day in a hot room, the contrast between body and background collapses and your easy segmentation stops working. An installation calibrated in winter can fail in summer.
Glass is opaque to LWIR. You cannot sense through a window or a vitrine, which rules out some mounting approaches that would work with a normal camera.
Reflective surfaces lie. Polished metal reflects thermal radiation and will produce ghost bodies.
It tells you nothing about appearance. Colour, clothing, gesture detail, facial expression — all unavailable. That’s the same property as the privacy advantage, seen from the other side.
For anyone building presence-aware work — particularly in dark rooms, public spaces or institutions with a privacy review — a self-contained thermal camera with on-device inference is a genuinely better primitive than the usual options, and this design is a concrete indication of where the price and integration are heading.