Why high-bay space is a hard problem
Point detectors work by waiting for combustion products to reach the sensor. In a warehouse with ten-metre ceilings, wide aisles and mechanical ventilation, that journey takes time, and smoke can be diluted or pushed sideways before it ever arrives. The result is a detection delay in exactly the environment where minutes decide whether an incident stays local.
Cameras have a different relationship with the problem. They see the volume of the room rather than a single point in it, so visible smoke in an aisle can be surfaced while it is still a plume rather than a layer.
What the system reports
Camera-based detection produces two event types, and the naming is deliberate:
- Possible Smoke — visual characteristics consistent with smoke were observed in a monitored area
- Possible Fire — visual characteristics consistent with flame were observed in a monitored area
Both are AI-assisted detections that require human confirmation. The system reports what the camera saw; a person verifies on site. An event is a prompt to look, and it is marked Manual Review Required until someone closes it.
This supplements certified systems, it does not replace them
This point needs stating plainly because it determines how the deployment is scoped. Camera-based detection is not a certified fire alarm system and does not satisfy fire safety regulation. It does not trigger suppression, and it must not be wired to replace an existing panel.
What it adds is an earlier prompt and a visual context that a point detector cannot provide. When an alarm does sound, responders already have a camera view of the origin area, which shortens the time spent locating the source.
Camera placement that actually works
Coverage planning matters more than model choice:
- Prioritise charging bays, battery rooms, packaging and shrink-wrap areas, and waste consolidation points
- Frame the aisle volume rather than the floor, since rising smoke is easier to see against a wall or rack face than against concrete
- Avoid views dominated by bright windows or roller doors, where backlighting washes out low-contrast smoke
- Keep at least one camera per fire compartment so an event can be located without ambiguity
Whether a given camera can carry a detection at all depends on its stream and framing, which Early Smoke Detection Using Existing IP Cameras sets out in detail.
Managing false positives honestly
Warehouses contain plenty of things that look like smoke. Diesel and LPG forklift exhaust, dust raised by pallet movement, steam from washdown areas, cold-store vapour at door openings, and low sun through skylights all produce candidate detections.
The workable response is environmental rather than algorithmic. Mask known exhaust and vapour regions, require a short persistence window before an event is raised, and calibrate per camera rather than per site. A detection that appears and disappears within two seconds is almost never worth a call-out.
The response workflow
A Possible Fire event should reach a person, not a log file. Define the path before go-live:
- Event raised with a clip, camera reference and zone
- Notification to the shift lead and the on-site safety contact
- Physical verification within a defined time window
- Outcome recorded as confirmed, explained or unclear
- Confirmed incidents follow the existing site emergency procedure without modification
Recording outcomes is what allows thresholds to be tuned and what demonstrates, during an audit, that the system is being operated rather than merely installed.
What to measure
Track median time from event to acknowledgement, share of events verified on site, share closed as explained by a known source, and the false-positive rate per camera per week. A single camera producing most of the noise is a placement problem, and it is usually fixable in an afternoon.
Safety events share their review workflow with security detections on the same cameras; Detect Suspicious Behaviour with AI Vision covers how that queue is kept short enough to stay credible.
