Fire and smoke on camera: how it works, and what it takes.

What camera-based fire and smoke detection is.

Camera-based fire and smoke detection uses video analytics to look for visible flame and smoke in a camera image. Instead of waiting for smoke to reach a ceiling detector or heat to reach a sprinkler head, the system looks at what the camera sees: a plume rising from a dumpster, flames in a pallet stack at the back of a yard, smoke drifting across a parking garage.

The appeal for security operators is obvious. The cameras are already there, and many of them cover places conventional detectors cannot: open yards, loading docks, rooftops, parking lots, solar farms and outdoor storage. Those are also the places where a guard is watching dozens of views and nobody is looking at the one where something just started.

How it works.

Older approaches used hand-built rules: look for flame-colored pixels, check for the characteristic flicker of fire, look for regions that become hazy and lose contrast the way smoke does. Those rules fired on sunsets, orange vests and taillights, and they missed faint grey smoke.

Modern systems use trained object detectors, the same family of models used for person and vehicle detection, trained on labelled images of smoke and flame. The detector draws boxes around candidate smoke and flame regions in each frame. That is only the first layer. A useful system adds:

• Temporal checks. Real smoke and fire persist and evolve. A single-frame hit is a possibility; the same class in overlapping regions over several seconds is much stronger evidence.
• Camera awareness. When the view changes, a feed drops or a camera cycles off screen, the evidence has to reset. A detection from one camera cannot confirm a detection on another.
• Context review. Some flame is fine: a grill, a fireplace, a welder, a fire on a TV screen. A scene-level review can explain what is visible so a guard can judge it quickly.
• A human decision. The guard looks at the evidence and follows post orders: call the fire department, call the client, send a patrol, or close it as controlled.

Why it is hard.

Smoke and fire are among the hardest things to detect reliably on security cameras. The lookalikes are everywhere:

• Steam and exhaust. Rooftop vents, kitchen exhaust, idling trucks on a cold morning and HVAC plumes all look like smoke.
• Fog, mist, dust and rain. Weather can wash out a whole scene in a smoke-like haze, especially under lights at night.
• Lights and glare. Headlights, sodium streetlights, low sun, sunsets and IR bloom all produce bright, warm-colored regions.
• Reflections. Wet pavement, glass facades and polished floors reflect lights that flicker and move.
• Orange things. Safety vests, traffic cones, signage and autumn leaves.
• Legitimate flame. Welding, grills, fire pits and candles.

Then there is image quality. On a monitoring desk, a camera is often a small tile on a busy dashboard, compressed and scaled down. Faint early smoke that is visible at full resolution can shrink to a few grey pixels. A model that works on clean still images can fall apart on real monitoring screens.

Camera-based detection complements certified fire alarm systems. It adds eyes on open areas and gives guards visual evidence; it does not stand in for smoke detectors, sprinklers or the life-safety systems your code requires.

Why it complements certified fire alarms.

Certified fire alarm systems are engineered, tested, installed and inspected to life-safety codes, and they detect things a camera cannot: heat, smoke inside closed rooms, fire behind walls or out of view. They also work in total darkness and when the camera network is down.

Cameras add different strengths: coverage of large open areas where smoke disperses before reaching any detector, visual confirmation for a guard, and a record of how an event started. The right model is layered: code-required fire alarm systems for life safety, and camera-based detection as an extra set of eyes that helps the security desk notice and verify sooner.

Our work: in active development.

Fire and smoke detection at Semaphor is in active development, with early access for pilot partners. We are building it the way it has to be built for a security desk, and we evaluate on real video, not demo stills.

Every run keeps model and dataset hashes, every prediction and every miss, backed by an automated test suite that covers class mapping, box matching, pane coordinates, gaps, camera changes and restarts. We hold these models to a service bar, not a demo bar: the first pilot checkpoint stays offline, and the next round expands the training data with more faint smoke and indoor hard negatives such as steam, glare and fire shown on screens. Every alert will follow the same rules as the rest of Semaphor: evidence first, guard decides.

Where Semaphor fits.

Semaphor already works from the monitoring screens your guards use today, with local person and vehicle detection, alert-first review and AI scene reports landing in the guard inbox beside the original frames. Fire and smoke alerts are designed to join that same loop as a separate hazard scan that does not depend on motion or people being present, with the guard verifying every alert against the evidence.

Operators with yards, warehouses, parking structures and outdoor storage are the best early partners. See what else is on the roadmap on what’s next, then join the pilot program or contact us for early access.

Fire and smoke detection questions.