The problem: by the time a fire is noticed, it is usually too late
Conventional fire detection waits for smoke to rise and physically reach a sensor. That works in enclosed, low-ceilinged, still-air rooms. But the areas that actually carry risk in a business rarely look like that:
- In high-ceilinged warehouses the fire has already grown by the time smoke reaches the sensor.
- In open yards, storage areas and construction sites smoke disperses and detectors never trigger.
- In greenhouses, farms and hangars humidity and dust make sensors unreliable.
- At night and at weekends nobody is on site — there is no human to see the fire either.
Meanwhile most facilities already have dozens of security cameras pointed straight at those risk areas. The only missing piece is something that watches them 24/7.
The solution: let the AI see what the camera sees
AI Kameram continuously analyses your live camera streams on a GPU server installed inside the facility. When flame or smoke appears, it is flagged as a detection within milliseconds. Verification layers then test whether that detection is genuinely a fire.
One deliberate design decision: the live stream never leaves your site. Analysis happens locally, and only the evidence frame and event record of a confirmed alarm are sent to the central panel. That protects your bandwidth and keeps outbound data minimal.
How fire detection works, step by step
1. Continuous image analysis
Each camera stream is scanned frame by frame by an object detection model trained to recognise flame and smoke. The model weighs not only colour but also shape, texture and motion.
2. Physical filters
Does the detected region match the physical characteristics of real flame? Size, aspect ratio and texture variance are checked. Flat, smooth bright spots — a lamp lens, a headlight, a reflector — are removed; the characteristic irregularity of real flame is required.
3. Static light filter
The system learns glares that repeat at the same spot. A fixed light fitting, a blinking LED or a surface the sun hits at a particular hour is progressively removed from the alarm path.
4. Persistence and growth checks
The signature of a real fire is temporal: flame flickers, the area grows, the behaviour is unstable. A single-frame flash does not raise an alarm; the event must persist consistently over a time window and/or grow.
5. Day / night scene mode
The system automatically recognises from the image that a camera has switched to night vision and is producing a black-and-white picture. Night decisions use different criteria: is it a bright white blob, is the texture too flat, is the shape too regular? This is why street lights and vehicle headlights are not mistaken for flame.
6. Second-stage visual verification
An event that passes every filter still goes through a second visual verification before becoming an alarm. The evidence crop is compared against the normal state of the scene. This stage is what gives the system the discipline not to cry wolf.
7. Zone and time rules
You draw which area of each camera is monitored, and attach time constraints to those zones. A welding bay can be excluded during working hours and monitored automatically after them; a canteen stove can be excluded entirely.
Why so many filters? Because the worst enemy of a fire alarm system is the false alarm. A system that goes off for nothing several times a day becomes, by week three, a notification nobody reads. The AI Kameram filter chain exists precisely to prevent that.
What happens when an alarm fires?
- Automated phone call — registered contacts are called and told the event type and location.
- SMS — a short alert arrives immediately.
- E-mail — with a screenshot, camera name, location and timestamp.
- Mobile push — instant notification to panel users.
- Guardhouse device — siren sounds, red light flashes, LCD shows the event.
At the same time the event video and evidence frame are archived, so they can be reviewed and reported afterwards from the panel.
Compared with a smoke detector
| Aspect | Smoke detector | AI Kameram (video analysis) |
|---|---|---|
| Trigger condition | Smoke physically reaching the sensor | Flame or smoke being seen by a camera |
| Open areas | Practically ineffective | Works |
| High ceilings | Delayed | No delay |
| Location detail | Zone level | Exact point in the camera image |
| Evidence | None | Screenshot and event video |
| Installation | Cabling and sensor mounting | Uses cameras you already have |
| Verification | None | Multi-layer filters + visual verification |
The two are not rivals: the detector infrastructure required by fire regulations stays in place, and AI Kameram is added on top as a visual early-warning layer.
Which cameras does it work with?
Any IP camera supporting ONVIF or RTSP, brand-independent. Analogue cameras connect through an NVR/DVR that exposes an RTSP stream. Thermal cameras are not required. It is enough that the camera has a clear view of the risk area and that its night vision works.
Frequently asked
How fast is camera-based fire detection?
Once flame or smoke enters the camera's field of view, detection and verification typically complete within seconds. Unlike smoke detectors there is no need to wait for smoke to rise to a ceiling sensor.
Do I need thermal cameras?
No. Standard IP cameras with night vision are sufficient. The system automatically detects when a camera switches to black-and-white night mode and applies night-specific filters.
Will welding, exhausts or stove flames cause false alarms?
Zone and time rules are defined for such sources; for example a welding bay can be excluded during working hours and monitored outside them. The static light filter and persistence/growth checks also remove repeating artificial sources.
Does it work outdoors and in high-ceilinged warehouses?
Yes, and these are exactly the scenarios where camera-based detection is strongest. Open yards, high-ceilinged warehouses, greenhouses and storage areas, where smoke detectors respond late or not at all, can be covered with video analysis.