Occupancy and presence sensing: PIR, mmWave radar and beyond

Walk into a meeting room and the lights come on. Settle in to read something quietly, and ten minutes later they switch off again, leaving you waving an arm at the ceiling to prove you are still there. That small, familiar annoyance is a sensing problem in miniature. The room could tell that something moved; it could not tell that a person was present.

Working out whether a space is in use — and, increasingly, how it is being used — is one of the quiet foundations of any building that claims to be intelligent. It decides when lights and heating run, whether an alarm should arm, and how a facilities team learns which desks and rooms are actually worth keeping. The trouble is that “is anyone there?” is a deceptively hard question, and the sensors that try to answer it disagree in ways that matter.

Four questions, not one

People use “occupancy sensing” loosely, but it hides four distinct jobs, each harder than the last.

• Motion detection asks only: is something moving? It is binary and cheap. Most automatic lighting is really motion-triggered, which is why it fails on a still person.
• Occupancy detection asks: is this space in use or vacant? In practice it is often motion detection plus a timer — the space is counted as occupied for, say, fifteen minutes after the last movement, then assumed empty.
• Presence detection asks: is a person here right now, even if they are perfectly still? Answering it means picking up signs of life that motion sensors miss, such as the rise and fall of a chest.
• People-counting asks: how many people are here? This is the basis of space-utilisation analytics, and it demands far more from a sensor than a simple yes or no.

A device that is excellent at one of these can be useless at another. A cheap motion sensor answers the first question well and the third one badly. Keeping the four apart is the single most useful habit when reading a sensor’s specification sheet, because vendors rarely draw the line for you.

The sensing technologies

Several physical methods can feed those questions, and none is best at everything.

Passive infrared (PIR). The PIR sensor is the workhorse behind most automatic lights and intruder alarms. It is passive — it emits nothing, only watching for changes in the infrared heat crossing its field of view as a warm body moves between the zones of its lens. That makes it cheap, low-power and private. It also means it detects change, not presence: sit still and, to a PIR, you effectively vanish.

Ultrasonic. An ultrasonic sensor emits sound above human hearing and listens for the frequency shift caused by movement. Because sound reflects around furniture and partitions, it can cover awkward, cluttered spaces and pick up smaller motions than PIR — a hand at a keyboard, for instance. The price is a tendency to false triggers from moving air, so a draught from an air vent can convince it the room is busy.

Microwave and millimetre-wave (mmWave) radar. Radar transmits a radio signal and reads what bounces back. Older microwave sensors simply flagged movement over a wide area, sometimes straight through a thin partition, which caused nuisance triggering. Modern mmWave radar, working in bands around 24 or 60 gigahertz, is a different proposition: it can measure how far away a reflection is and detect motion on the scale of a breathing chest. That lets it confirm a still person is present, and even estimate how many targets are in a room, without forming a recognisable image of anyone.

Thermal and thermopile. A thermopile array is, in effect, a very low-resolution thermal camera — a grid of a few dozen to a few hundred pixels reading emitted heat. Because a warm body radiates whether or not it moves, a thermal array can hold a stationary person and count several at once, while its coarse resolution means you cannot identify individuals from the output.

Camera and computer vision (CV). A conventional camera paired with computer vision is the most capable option by a wide margin. It can count people, track movement, tell a person from a coat stand, and read queue lengths. It is also the most intrusive, the most demanding to process, and the one that turns a sensing question into a surveillance question.

Carbon dioxide as an indirect proxy. People exhale carbon dioxide (CO2), so a room’s CO2 level rises as it fills. Outdoor air sits around 400 parts per million (ppm); an occupied, poorly ventilated room climbs well above that. Measuring CO2 says nothing about who or where, and it lags real occupancy by many minutes as the gas accumulates and disperses, but it is a cheap, genuinely private way to drive ventilation to match how busy a space is.

Wi-Fi and Bluetooth (BLE) counting. Since most people carry a phone, counting the devices probing for Wi-Fi or Bluetooth Low Energy (BLE) gives a rough headcount across a large area such as a shop floor or a concourse. It is convenient and needs no new sensors, but the numbers are soft: phones now randomise their identifiers to resist tracking, some people carry two devices and others none, and a switched-off handset is invisible.

Comparing the options

Technology	Best at	Strengths	Weaknesses	Privacy exposure
PIR	Motion detection	Cheap, low-power, emits nothing	Blind to still people; needs line of sight	Very low
Ultrasonic	Occupancy in cluttered rooms	Sees around obstacles; catches small motions	False triggers from moving air	Low
Microwave / mmWave radar	Presence detection, some counting	Detects breathing-scale motion; can gauge range	Can over-reach through partitions; costlier	Low to moderate
Thermal / thermopile	Presence and small-group counting	Holds stationary people; counts a few at once	Limited range; coarse detail	Low to moderate
Camera + computer vision	People-counting and tracking	Highest accuracy; rich detail	Processing-heavy; most intrusive	High
CO2	Ventilation control	Cheap; whole-room; no personal data	Slow to respond; no location or count	Very low
Wi-Fi / BLE	Rough headcount over wide areas	Uses existing infrastructure	Imprecise; defeated by MAC randomisation	Moderate

The empty-room problem

The clearest illustration of why any of this matters is the lighting that switches off on a motionless person. A PIR sees you cross the room, starts its countdown, and — while you sit reading, typing gently, or listening in a call — sees nothing worth resetting the timer for. When the countdown ends, the lights go out on someone who never left. Cranking the timeout up to fifteen or twenty minutes hides the fault but wastes the energy the sensor was meant to save.

This is the gap presence detection exists to close. mmWave radar can register the sub-millimetre chest movement of breathing, so a person sitting perfectly still still reads as present. The room stays lit while it is genuinely occupied and goes dark promptly once it is empty — the opposite of the usual trade-off between comfort and waste. The same ability to detect breathing-scale motion is what makes radar attractive for ambient assisted living, where a sensor that can tell “still present” from “collapsed and not moving” is doing safety-critical work without a camera in the room. In many designs, combining a fast, cheap PIR with a confirming radar or thermal channel gives quicker reactions and fewer false absences than either sensor alone.

What it is used for

Occupancy and presence data earn their keep in three broad areas.

1. Lighting and HVAC control. This is the everyday case: lights, heating, ventilation and air conditioning (HVAC) that follow real use rather than a fixed timetable. A building management system (BMS) that knows which rooms are occupied can dim empty corridors and stop conditioning air for nobody, which is a large share of what people mean when they call a building efficient. Where this fits into the wider picture is covered in our guide to what makes a building smart.
2. Security. Presence sensing arms and disarms zones, distinguishes an out-of-hours intruder from a cleaner on shift, and flags a room that should be empty but is not.
3. Space-utilisation analytics. Aggregate counts over weeks answer the questions that decide leases and layouts: which meeting rooms sit booked-but-empty, how full the office really is on a Wednesday, whether a floor could be handed back. Here the appeal of anonymous counting — radar or thermal rather than cameras — is that it measures patterns without recording people.

The privacy gradient

Run the technologies from PIR to camera and you trace a steady climb in what a sensor can know about you. A PIR records only that heat moved; there is essentially nothing personal to leak. Radar and thermal arrays know your position and rough number without an identifiable image. A camera with computer vision can, in principle, recognise faces, log comings and goings, and infer far more than whether the lights should be on.

The useful rule is to choose the least revealing sensor that answers your actual question. If you only need to switch lights, you do not need to see faces; a camera used for a job a radar could do is collecting risk for no extra benefit. Where richer sensing is genuinely warranted, on-device processing helps — a thermal or radar module that outputs “three people” rather than raw frames keeps the sensitive data from ever leaving the room. These choices, and the governance around them, are the subject of our guide to privacy in intelligent environments.

Choosing well

There is no universal winner here, only sensible fits. For a stairwell light, a PIR is the right answer and anything fancier is wasted money. For a focus room where people sit still, radar or thermal sensing earns its place. For ventilation, a CO2 monitor quietly does a job the motion sensors cannot. And for people-counting, the honest question is not which sensor is most accurate but how much detail you are entitled to collect to get the number.

The direction of travel is towards presence over mere motion, and towards sensors that answer usefully while revealing as little as possible about the person who prompted it. A room that stays lit because it can tell you are breathing, without ever needing to see your face, is a fair example of what good sensing should feel like: helpful, and easy to forget is there at all.