Edge computer vision is the quiet inversion of a decade of cloud thinking: instead of shipping video to a data centre so software can look at it, the software moves into the camera and the video never travels at all. That one architectural decision changes what a camera is. It stops being a recording device and becomes a sensor, an instrument that emits numbers, counts, densities, dwell times, and forgets each frame the moment it has been measured. This post is an educational tour of computer vision on the edge: what it means, how an edge AI platform for video and sensor analytics is actually put together, and what the numbers unlock in retail, events and the built environment.
01 · WHAT COMPUTER VISION ON THE EDGE ACTUALLY MEANSWhat computer vision on the edge actually means
Edge computing is the general principle of processing data close to where it is produced rather than in a centralised cloud. Computer vision on the edge applies that principle to cameras: a small computer, usually an embedded system-on-chip mounted in or beside the camera, runs the detection model on-device. The pixels are analysed where they are captured.
The contrast with conventional video analytics is stark. A cloud pipeline streams every frame across the network, stores it, then analyses it in a data centre. An edge pipeline analyses the frame in place, keeps the measurement and discards the frame. Same computer vision, opposite data gravity.
Three properties fall out of this architecture, and they explain almost every real-world deployment decision:
- Privacy is structural. Raw frames never leave the device, so there is no footage to store, leak, subpoena or misuse. The privacy promise is enforced by the architecture, not by a policy document.
- Latency is sub-second. The measurement happens where the light lands. There is no round trip to a data centre, so a count or an alert reflects the room as it is right now.
- Bandwidth collapses. A single 1080p camera streams several megabits per second. An occupancy update is a few hundred bytes. That difference decides whether a deployment of fifty sensors needs a network upgrade or a spare switch port.
02 · WHY THE FOOTAGE NEVER LEAVES THE DEVICEWhy the footage never leaves the device
The obvious question about any camera-based system is the uncomfortable one: who is watching? With edge computer vision the honest answer is nobody, because there is nothing to watch. The model detects people as anonymous shapes, converts them to coordinates and counts, and the frame is gone from memory within milliseconds. No video is stored. No video is transmitted. No faces are recognised.
This matters legally as well as ethically. Under the Australian Privacy Principles, collecting identifiable footage of individuals drags an organisation into collection, storage, retention and disclosure obligations. A sensor that never produces identifiable data in the first place keeps whole categories of risk off the register. We call this privacy by design, and it is the reason a university, a council and a retailer can all say yes to the same architecture.
The camera becomes a sensor. The output is a number, not a picture, and you cannot leak a picture you never kept.
03 · ANATOMY OF AN EDGE AI PLATFORM FOR VIDEO AND SENSOR ANALYTICSAnatomy of an edge AI platform for video and sensor analytics
A production system is more than a clever camera. An edge AI platform for video and sensor analytics has four layers, and it is worth understanding each because procurement questions usually map to exactly one of them.
1. Capture and inference. The sensor observes a zone and runs a person-detection model on every frame, on-device. Output per frame: how many people, roughly where. The frame is then discarded. Calibration per site matters here, because a lens mounted at three metres over a cafe sees very differently from one at eight metres over a concourse. Our counting accuracy deep dive covers why we publish confidence intervals rather than a single marketing number.
2. Aggregation. Counts stream off the device over MQTT or HTTPS, a few bytes at a time, into a time-series store. This is where instantaneous detections become the metrics people actually use: occupancy per zone, density per square metre, dwell time distributions, queue lengths, arrival and departure rates.
3. Sensor fusion. Vision is one input. The same platform ingests non-visual signals, CO₂ and air quality being the most useful pairing, because CO₂ lags occupancy in a predictable way. Counts explain the CO₂ curve; the CO₂ curve validates the counts. Fused together they describe how a space is really being used, which is the substance of occupancy analytics.
4. Delivery. Dashboards for humans, an occupancy API and webhooks for machines. More on that below, because this layer is where the system stops being interesting and starts being useful.
04 · RETAIL DWELL: THE NUMBER STORES ACTUALLY ACT ONRetail dwell: the number stores actually act on
Footfall tells a retailer how many people came in. Retail dwell tells them what happened next, and it is the more commercially interesting number. Dwell time is how long a shopper spends in a zone: at a display table, in an aisle, near a fitting room, at the counter.
Measured anonymously and per zone, dwell answers operational questions that footfall cannot:
- Which window display converts passers-by into browsers, and which one gets walked past?
- Where do shoppers linger without buying, a signal of interest the range is not converting?
- How long is the queue tolerance before people abandon, by hour and by day?
- Did the new layout actually change how people move, or just how the plan looks?
None of this requires knowing who anyone is. A dwell distribution is a statistic about a zone, not a profile of a person, which is why edge computer vision fits retail environments where a facial-recognition trial would be a headline. The broader retail picture lives on our retail industry page.
05 · EVENTS OCCUPANCY MONITORING AT CROWD SCALEEvents occupancy monitoring at crowd scale
The same architecture stretched across a stadium concourse or a festival ground becomes a safety instrument. Events occupancy monitoring is less about a total headcount and more about live density and flow per zone: how full is this concourse, how fast is it filling, and is the gate keeping up with the queue?
Crowd safety practice treats density in persons per square metre as the variable that predicts trouble before it is visible, and it has to be read live to be useful. That is only possible when inference happens at the edge, because streaming forty camera feeds to a cloud service and waiting for answers adds exactly the latency a control room cannot afford. We wrote a full deep dive on concourse monitoring and outdoor venue occupancy analytics, including the density thresholds the crowd-science literature uses, and the events and stadiums page covers the operational side.
06 · FROM OCCUPANCY API TO ACTION: WEBHOOKS, ALERTS AND ESCALATIONFrom occupancy API to action: webhooks, alerts and escalation
A measurement that stays on a dashboard is a curiosity. The delivery layer is where edge computer vision earns its keep, and it is deliberately boring technology: a REST occupancy API you can poll, and webhooks that push.
A typical occupancy API response is small enough to read whole:
{
"zone": "concourse-north",
"occupancy": 214,
"density_ppm2": 2.4,
"dwell_p50_s": 96,
"trend": "rising",
"ts": "2026-07-06T09:41:07+10:00"
}
Because the payload is just structured numbers, it plugs into whatever an operations team already runs:
- Incident escalation. A density threshold crossing becomes a PagerDuty incident, so a crowding event pages the duty manager the same way a server outage pages an engineer, with the same acknowledgement and escalation discipline.
- Messaging. The same threshold can post to Slack or Teams for the softer cases, a queue creeping past its service target, a floor filling faster than staffing assumed.
- Building systems. Occupancy feeds the BMS so HVAC follows real usage instead of a worst-case schedule.
- Data platforms. Counts land in Snowflake or a warehouse next to sales and rostering data, where dwell against conversion becomes an analysis rather than a hunch.
The full catalogue lives on the integrations page and the alerting model on the alerts page. The design principle is simple: the platform should meet the tools an organisation already trusts, not demand a new pane of glass.
07 · EDGE OR CLOUD: AN HONEST COMPARISONEdge or cloud: an honest comparison
Cloud video analytics is not wrong, it is a different set of trade-offs. The honest comparison looks like this:
| Dimension | Edge computer vision | Cloud video analytics |
|---|---|---|
| Raw video | Never leaves the device | Streamed and usually stored |
| Privacy posture | Structural, no footage exists | Policy-based, footage must be protected |
| Latency | Sub-second, local | Seconds or more, network dependent |
| Bandwidth per camera | Bytes per update | Megabits per second, continuously |
| Offline behaviour | Keeps counting, syncs later | Blind while the link is down |
| Retrospective review | Not possible, nothing stored | Possible, that is the point of storing |
| Typical fit | Live measurement of spaces | Forensic review, security recording |
The last two rows are the honest part. If an organisation genuinely needs to scrub back through footage, edge computer vision is the wrong tool, because there is deliberately nothing to scrub. If the need is live measurement of how spaces are used, streaming and storing video is cost and risk with no compensating benefit.
08 · WHERE TO STARTWhere to start
The pattern we recommend is unglamorous: one space, one question, a few weeks of data. Pick the concourse that always feels crowded, the floor with the disputed utilisation, the display wall the merchandising team argues about, and measure it. The first dataset usually settles an argument someone has been having for years, and it does so without a single frame of video existing anywhere.
If you have a space like that, book a walkthrough and we will scope what a sensor would see, or start with the edge AI technology page if you want the architecture in more depth first.