TIER 1
Reference SoC
Jetson-class sensor with our pre-flashed model. Recommended for net-new pilots. PoE, ceiling or wall mount. From AUD $390 per sensor.
01 · Platform · Occupancy sensor
The occupancy sensor that counts people without watching them.
Most occupancy sensors force a trade: the accuracy of a camera or the privacy of a presence detector. Ours refuses the trade. It has the optical accuracy of a camera and the privacy of a motion sensor, because the frame is processed on the device and discarded before anything leaves the room. No video stored, no video transmitted, no faces recognised.
02 · Definition
What a privacy occupancy sensor actually measures.
An occupancy sensor reports how a space is used. The cheap end of the market reports a single bit, occupied or empty. We report the numbers a building runs on: live counts per zone, density against capacity, and dwell time. All of it is anonymous. The sensor produces numbers, not records of people, so there is nothing in the output that could identify anyone.
Counts
People in, people out, people now, per zone. The base layer every other metric is built on.
Density
Counts measured against the capacity of the zone, so a number becomes a signal you can act on.
Dwell
How long people stay, anonymised. Useful for utilisation and merchandising without tracking anyone.
03 · How it works
Camera, edge SoC, counts. The frame never leaves the device.
This is the whole pipeline. An image sensor feeds an on-device system-on-chip; the chip runs the model, emits a count and discards the frame. Read the full architecture on the on-device edge AI page.
01 · Camera
Wide-angle image sensor
Standard CMOS. Outputs raw frames to the SoC over MIPI. Those frames never reach a network interface.
02 · Edge SoC
On-device inference
A Jetson-class chip runs the counting model, emits a count vector and density grid, then deletes the frame. There is no face model anywhere in the pipeline.
03 · Counts
Anonymous JSON out
Counts, deltas and threshold events, encrypted in transit and stored aggregated only. No raw video ever existed to upload.
camera to counts in < 60 ms, end to end
04 · Compare
People counting sensor types, compared honestly.
Every modality earns its place somewhere. The point of this table is not that the others are bad, it is that only on-device optical wins on accuracy and privacy at the same time.
| Modality | Count accuracy | Privacy | Counts vs presence | Relative cost | Best for |
|---|---|---|---|---|---|
| PIR (passive infrared) | Low | High (no imaging) | Presence only | Low | Lighting and switch-off triggers |
| Wi-Fi / BLE counting | Low to medium | Weak (device re-ID risk) | Rough count | Low | Coarse footfall where devices are carried |
| mmWave / radar | Medium | High (no imaging) | Count, limited density | Medium | Small rooms and presence with motion |
| Thermal imaging | Medium | High (low resolution) | Count | Medium to high | Doorway counting in low light |
| Time-of-flight / depth | Medium to high | High (no RGB stored) | Count, some density | High | Single entrances and tight thresholds |
| Occivar edge-optical (on-device) | 97.8% vs ground truth | High (frame discarded on device) | Count, density and dwell | Medium | Accurate, anonymous occupancy across whole zones |
Accuracy figures for non-optical modalities are general industry ranges; the 97.8% optical figure is measured against ground truth in well-lit indoor conditions.
05 · Specs
The numbers, and the three ways to deploy it.
Inference latency
< 60 ms
Camera to counts, per frame, p99
Counting accuracy
97.8%
vs ground truth, well-lit indoor
Video frames stored
0
On device or in cloud
Bandwidth per sensor
≈ 200 B/s
Counts only, JSON over HTTPS
Power and mounting
PoE
Ceiling or wall, single cable
Reference price
AUD $390
From, per sensor (Tier 1)
TIER 2
Bring your own edge
Already have x86 or ARM compute? We ship the runtime. Linux only, containerised, with the same on-device frame-discard guarantee.
TIER 3
Existing CCTV (read-only)
Where compliance permits, run the model against an existing camera feed on a co-located edge box. The frame-discard contract still holds.
06 · Privacy
No biometrics, no demographics, no exceptions.
The sensor has no face model, no gait model and no demographic classifier, so there is nothing to leak, subpoena or breach. It is aligned with the Australian Privacy Act and GDPR by construction rather than by promise. See the full privacy-by-design architecture for how the guarantee is enforced in hardware and code.
07 · Integrations
The counts land where your building already works.
Anonymous counts flow into your building management system, Slack, Teams, Snowflake, a REST endpoint or an MQTT broker over HTTPS. Pair the sensor with live occupancy analytics for dashboards and thresholds, and see the full list of BMS and data integrations.
08 · FAQ
Occupancy sensor questions, answered plainly.
Still deciding between sensor types? The blog goes deeper on whether occupancy sensors use cameras.
Do occupancy sensors use cameras?
Some do and some do not. PIR, mmWave radar and time-of-flight sensors use no camera but can only report presence or a rough count. Optical sensors use an image sensor to count accurately. An Occivar occupancy sensor uses an image sensor, but the frame is processed on the device and discarded in milliseconds, so no video is ever stored or transmitted. You get optical accuracy without optical privacy risk.
How accurate is a camera-free occupancy sensor?
It depends on the modality. PIR and Wi-Fi or BLE counting are typically within tens of per cent of ground truth and degrade in busy or open spaces. On-device optical counting measures 97.8% against ground truth in well-lit indoor conditions, because it counts people directly rather than inferring them from motion or device signals.
Is an occupancy sensor compliant with the Australian Privacy Act and GDPR?
A sensor that stores or transmits identifiable video collects personal information and pulls you into the full obligations of the Australian Privacy Act and GDPR. An Occivar sensor never stores or transmits a frame and holds no biometrics, so the data leaving the room is anonymous counts. Privacy is enforced by the architecture rather than by a retention policy you have to trust.
What is the difference between an occupancy sensor and a PIR, mmWave or Wi-Fi sensor?
PIR detects motion and reports presence, not a count. mmWave radar handles small rooms well but struggles with density and dwell. Wi-Fi or BLE counting infers people from device signals, which MAC-address randomisation breaks and which carries re-identification risk. An on-device optical occupancy sensor counts people directly and anonymously, giving accurate counts, density and dwell without holding any identifier.
How are Occivar occupancy sensors mounted and powered?
The reference sensor is ceiling or wall mounted and powered over Ethernet (PoE), so a single cable handles power and the count data. It is hardware agnostic: you can run our reference SoC, bring your own edge compute, or, where compliance permits, run the model against an existing camera feed on a co-located edge box.
What data leaves the sensor?
Only anonymous JSON: counts, density and threshold events, at roughly 200 bytes per second over HTTPS. No video, no images, no faces and no PII ever leave the device. There is no raw footage to upload because the frame is discarded the moment the count is produced.
09 · Put it on your ceiling