Reusing the estate you already paid for
Most sites considering visual smoke detection already have the hardware. A typical facility runs dozens of IP cameras on an existing recorder, and the practical question is not whether to buy new equipment but which of the current views can carry a Possible Smoke detection reliably.
The answer is usually some of them. Being specific about which ones, before a pilot starts, is the difference between a deployment that gets extended and one that quietly stops being trusted.
Stream requirements in plain terms
Smoke is a low-contrast, slowly evolving, semi-transparent phenomenon. That makes it unusually sensitive to stream quality:
- Resolution: 1080p is generally sufficient for a plume within twenty-five metres; 720p narrows that considerably
- Frame rate: 10–15 fps is enough, since smoke develops over seconds rather than frames
- Bitrate: this matters more than resolution, because aggressive compression smooths away exactly the texture the detection depends on
- Protocol: a standard RTSP or ONVIF sub-stream is normally the cleanest integration point
- Clock accuracy: NTP sync across cameras, so events line up on one timeline
If a camera can only offer a heavily compressed sub-stream, it is better to exclude it than to include it and accept a detection that misses.
What a view needs to see
Three placement properties predict success better than anything else:
- A contrasting background. Smoke against a dark rack face or a painted wall is visible; smoke against a bright skylight or a white ceiling frequently is not.
- Vertical headroom in frame. A plume rises, so a view that includes several metres above the likely ignition area detects earlier than one cropped to floor level.
- Stable lighting. Cameras facing roller doors, windows or vehicle headlights see rapid global brightness changes that resemble the onset of smoke.
Cameras that fail these tests still have a role. They provide context and timeline evidence once an event is raised elsewhere, which is useful even when they cannot originate a detection.
Edge processing and bandwidth
Running inference at the edge, close to the cameras, avoids sending continuous video to a central service. Only events and short evidence clips leave the site, which keeps bandwidth predictable and means detection continues during a connectivity outage, with events buffered and reconciled once the link returns — the mechanism described in How Offline Recovery Works.
Latency budget is worth stating explicitly. From the first visible smoke to a notification on someone's phone, a realistic target is under fifteen seconds, most of which is the persistence window that suppresses momentary artefacts.
Known sources of false detections
Being candid about these upfront sets the right expectations:
- Vehicle and forklift exhaust in loading areas
- Steam from washdown, kitchens or cold-store door openings
- Dust raised by sweeping, pallet movement or nearby construction
- Fog, low cloud and heavy rain on external cameras
- Direct sun and moving shadows through skylights
Most are handled by masking the source region, requiring persistence before an event is raised, and tuning each camera individually. Some are seasonal and are best handled by scheduling sensitivity rather than by chasing a universal threshold.
Scope and limits
Visual detection sees what the camera sees. Smoke inside a sealed enclosure, behind stacked racking, or in an unmonitored room will not be detected, and no amount of tuning changes that. This is a supplementary early-warning layer that works alongside certified detection and alarm systems, not a substitute for them, and every Possible Smoke event remains Manual Review Required until a person verifies it on site.
A sensible pilot
Choose four to six cameras covering genuinely elevated-risk areas, run for three weeks, log every event with an outcome, and review the per-camera false-positive rate at the end. Cameras that produce noise are re-aimed or excluded, and the ones that stay become the baseline for a wider rollout. For coverage planning and the response workflow that sits behind these events in a large facility, see AI Fire & Smoke Detection for Warehouses.
