Camera-Based Fire Detection: AI That Sees Before the Detector

In high-ceilinged and open areas, camera-based fire detection sees flames and smoke minutes before a detector does. How it works, camera choice and its limits.

Translated from the Turkish original · Türkçe aslı

Camera-based fire detection is an early warning method that uses AI to look for the visual signature of flames and smoke in footage from existing security cameras. The model evaluates colour distribution, flicker texture and the spreading geometry of smoke together over consecutive frames; it does not wait for the smoke to reach the detector on the ceiling. In high-ceilinged and open areas, this means a warning with an image that arrives minutes before the detector.

A smoke detector waits for the smoke to reach the ceiling; in open and high-ceilinged sites this can take minutes. Camera-based fire detection, on the other hand, sees flames and smoke directly at the point where they form, as long as they are within the field of view. A confirmed event is reported to the person responsible with its image and camera location; the person receiving the alert does not first have to go to the site to verify the event.

How does camera-based fire detection work?

Detection runs on two separate signals. For flames, the colour histogram and flicker texture are read together: a real flame's edge undulates irregularly, its brightness changes from frame to frame and it grows upwards from its source. For smoke, the geometry of the motion matters: smoke rises, widens, softens the edges of the scene behind it and lowers the contrast. Steam also rises, but with a different texture and it disperses faster; dust and vehicle exhaust mostly move close to the ground and horizontally.

A single frame does not trigger an alarm. A candidate region becomes an event only when it shows consistent behaviour over a window of a few seconds; this temporal confirmation prevents a momentary flash or reflection from turning into an alarm. The system works with existing cameras that provide an RTSP or ONVIF stream, also via an NVR or VMS, and no additional sensors are installed on site.

Why can a camera see earlier than a smoke detector?

A point-type smoke detector waits for smoke to rise to the ceiling and enter the sensing chamber at sufficient density. In a low-ceilinged room this path is short. In a high-bay warehouse or an open site, however, smoke cools as it rises, stratifies, is dispersed by air currents and reaches the detector late; in an open area it may never reach it. Because the cost of a fire grows rapidly with time, the metric to track is not the number of alarms but the detection-to-response delay.

In the sample dashboard, the difference is about 47 seconds in an enclosed area and about 2.4 minutes under a high ceiling; that is why the phrase “minutes before” applies to high-ceilinged and open areas. This is also where the method's limit lies: a camera only detects what it sees. It cannot see a fire that starts inside a suspended ceiling, in a cable duct, in an electrical panel or behind a rack and has not yet produced visible smoke. The flame/smoke detection accuracy published on our solution page is 95%, and the time from detection to notification is under 10 seconds; these values apply to events within the line of sight.

How are false alarms reduced in flame and smoke detection?

Welding light, sunlight reflecting off glass or metal, orange warning lamps and vehicle headlights can be confused with flames; process steam, dust and exhaust with smoke. That is why the model looks not at a single frame but at behaviour over time: a real flame flickers irregularly and grows, while a reflection stays still or moves with the light source; steam disappears quickly, while smoke accumulates. The second layer is the rule engine: known hot spots such as a welding station, a furnace mouth or a cutting bench are defined as separate zones and monitored with a different threshold or time rule.

The most common sources of false alarms on site are steam, welding sparks and vehicle headlights; these are eliminated one by one during the pilot, and the threshold is updated until the rate stabilises. In the sample pilot curve on our solution page, the monthly number of false alarms falls by −72% through this elimination process; the value at your site is measured in the pilot. In fire detection the priority is sensitivity: accepting a few extra alerts is better than missing an event. We explained this balance in the confidence score, precision and recall article, and zone and duration rules in the reducing false alarms article.

What camera, angle and distance are needed for fire detection?

A standard visible-light camera is sufficient; a thermal camera is optional. What matters is not the camera's brand but how large it sees the flame source in the frame: the flame source should cover at least 20 pixels in the image, and under this condition effective detection is possible up to 60 m. The camera must be fixed; with a continuously patrolling PTZ camera the background changes all the time and the monitored area periodically leaves the frame.

Lighting conditions affect flames and smoke differently. Flame detection works with higher accuracy at night, because the flame's contrast increases in a dark scene. Smoke, on the other hand, needs light to be visible; when the camera switches to infrared mode, colour information is lost and the trace of smoke weakens. At these points, lighting or a thermal camera is evaluated during the site survey. We covered the choice of angle and position in detail in the camera placement article.

How does camera analytics differ from smoke detectors and thermal cameras?

These technologies are not rivals but answers to different risk points. Camera analytics does not replace the detector: the fire detection system stays in place as required by regulations and insurance, and video analytics is an image-verified early warning layer added in front of it. At points where a thermal camera is present, flame analytics can be cross-verified with thermal data; we described the hybrid setup in the thermal cameras and AI article.

Which areas should be prioritised, and where is it not suitable?

Camera analytics makes the biggest difference in volumes where the detector is structurally late. In high-bay warehouses smoke reaches the detector minutes later; escape routes, flammable material zones and stacking rules can be inspected with the same cameras under warehouse and hazardous material monitoring, with details in the warehouse safety article. Open-yard piles at recycling facilities carry a risk of spontaneous combustion, and detectors do not work in open areas; adding a thermal camera makes sense to catch the heating before flames appear. In production, welding and cutting stations are typical points for spark-induced incipient fires; in enclosed car parks, the first seconds of a vehicle fire are critical.

The places where it is not suitable are also clear. Narrow volumes without a line of sight, the inside of electrical panels and suspended ceiling voids are outside the scope of this method. Process areas that continuously produce flames, smoke or steam, such as casting, furnace or drying lines, can only be monitored with careful zone and time rules; sometimes it is better to leave these areas out of analysis and to the existing detectors.

What happens when a fire alarm comes in, and which systems does it connect to?

A confirmed event becomes a notification in under 10 seconds. The notification includes an image of the event and the camera location; escalation is tiered: the first alert goes to security, and if intensity increases it is automatically escalated to the next level. Outputs can be connected to the fire alarm panel, to a siren via an IoT relay and to phone notifications; because events are published over REST API and MQTT, they are also transferred to the VMS or building management system. Every event is archived as a timestamped clip; it serves as a record both in audits and in post-incident review.

It is recommended that hard-to-reverse actions such as automatic suppression and evacuation remain tied to the certified fire detection system. Article 11 of Law No. 6331 (Türkiye's Occupational Health and Safety Law) obliges the employer to assess emergencies in advance, prepare an emergency plan and assign personnel for firefighting; archived clips provide a documentable record for drills and post-incident review. In the default setup, footage is processed on an edge server inside the facility, raw video is not sent to the cloud (processing can also be done in the cloud if preferred) and facial recognition is not used; we covered the KVKK (Türkiye's Personal Data Protection Law) side in the KVKK-compliant video analytics article.

What are the common mistakes on site?

• Putting the system in place of the existing fire detection infrastructure; camera analytics is a complementary, accelerating layer.
• Choosing cameras based on the existing layout rather than the risk points; a camera that is too far away cannot see the flame at the required pixel size, and an angle that sees behind the racks often does not exist at all.
• Switching on all rules on day one without defining known hot spots and process steam; unnecessary alerts end with the system being muted.
• Skipping the night test; smoke performance must be verified separately on cameras that switch to infrared mode.
• Not writing down the notification chain; an alert sent without deciding who responds and in what order does not speed up the response.

Where should you start with camera-based fire detection?

The starting point is a risk map: areas with a high fire load, where detectors are late and where human supervision is limited are identified. For ready-made solutions, installation is typically completed within a week, followed by a 30-day pilot. At the start of the pilot, it is written down which camera will monitor which volume, what counts as flame or smoke, and to whom the alert will go; every alert the system produces is reviewed with the site team and marked as a real event, steam or a spark. In facilities where controlled smoke tests are possible, detection delay is also measured with these tests. To set up the success criterion, see the designing the right pilot article.

At the end of the pilot you get a suitability report per camera, a table of events caught and missed per scenario, and calibrated thresholds; the same report also shows where cameras need to be added or which angle needs to be changed. CX Teknoloji assesses the suitability of existing cameras for flame and smoke analysis free of charge before installation.

Frequently asked questions

Does camera-based fire detection replace the smoke detector?

No. Smoke and heat detectors are the basic layer, connected to the fire alarm panel, that stays in place as required by regulations and insurance. Camera analytics is an early warning layer added in front of it: in high-ceilinged and open areas where the detector is late, it catches flames and smoke in the image and adds the event image and camera location to the notification. It is recommended that actions such as automatic suppression and evacuation remain tied to the certified system.

Can fire detection be done with our existing CCTV cameras?

In most cases, yes. The system works with IP cameras that provide an RTSP or ONVIF stream, also via an NVR or VMS; a standard visible-light camera is sufficient, and a thermal camera is optional. The conditions are that the flame source covers at least 20 pixels in the frame and that the camera is fixed; under these conditions, effective detection is possible up to 60 m. Your camera inventory is assessed free of charge before installation.

How accurate and fast is camera-based fire detection?

The flame/smoke detection accuracy published on our solution page is 95%, and the time from detection to notification is under 10 seconds. Results vary with the camera angle, the size of the flame in the frame, lighting conditions and misleading sources on site; these values apply to events within the line of sight. A single frame does not trigger an alarm, and thresholds are calibrated on site during the pilot.

Do welding sparks, steam or vehicle headlights cause false alarms?

These sources are the most common misleaders on site. The model distinguishes them by their behaviour over time: a real flame flickers irregularly and grows, a reflection stays still, and steam disperses quickly. Known hot spots such as welding stations are monitored with a separate zone and threshold. The remaining cases are eliminated during the pilot, and thresholds are updated until the rate stabilises.

Does camera-based fire detection work at night?

Flame detection works with higher accuracy at night, because the flame's contrast increases in the dark. The situation is different for smoke: it needs light to be visible, and when the camera switches to infrared mode, colour information is lost. In critical areas that must also be monitored at night, lighting should be maintained, night footage should be tested separately in the pilot, or the layer should be completed with a thermal camera.

Where is the footage processed, and what about KVKK?

In the default setup, footage is processed on an edge server inside the facility and raw video is not sent to the cloud; processing can also be done in the cloud if preferred. Fire detection focuses on flames and smoke, not people; facial recognition is not used and event clips are automatically deleted at the end of the defined retention period. Obligations such as the privacy notice and the retention period remain with the facility as data controller; this information does not replace legal advice.