Computer vision · Smoke and fire
Smoke, seenearlier.
Visual smoke and fire detection is the computer vision technique that recognises the patterns typical of smoke and flames in camera images. We build systems that watch areas that are hard to cover with conventional detectors, such as yards, high-bay warehouses or open spaces, and send an alert with the related image to the people who can check. It is an additional warning layer and does not replace certified fire detection and alarm systems.
Events
- 14:08:51Possible smoke · aisle C4
- 14:08:53Confirmed on consecutive frames
- 14:08:53Alert sent to the shift lead
- 14:08:54Certified system stays primary
Some areas remain hard to cover
Point smoke detectors work well in ordinary rooms, but in very high spaces, in yards or outdoors, smoke can disperse before it reaches them. Cameras already capture these areas, yet nobody can watch every image all the time.
- In open or very high spaces, smoke reaches a conventional detector late or not at all
- Existing cameras film the sensitive areas, but the images are only looked at occasionally
- A fire starting outside working hours may only be noticed once it is visible from afar
- The control room operator receives many images at once and has no signal showing where to look first
Who it is for
- Warehouses and logistics hubs with high or large spaces, where smoke reaches ceiling detectors late
- Outdoor storage yards and sites holding waste or combustible materials
- Manufacturing sites that already have cameras and want an extra layer of vigilance over sensitive areas
- Security control rooms that watch many cameras and need a priority visual alert
Step by step
How the system spots smoke, step by step
The animation follows the start of a fire in a warehouse aisle, from the camera image to the person who checks it. The five steps below show what happens at each stage.
Events
Events
- 14:08:51Possible smoke · aisle C4
Events
- 14:08:51Possible smoke · aisle C4
- 14:08:53Confirmed on consecutive frames
Events
- 14:08:51Possible smoke · aisle C4
- 14:08:53Confirmed on consecutive frames
- 14:08:53Alert sent to the shift lead
- 14:08:54Certified system stays primary
Events
- 14:08:51Possible smoke · aisle C4
- 14:08:53Confirmed on consecutive frames
- 14:08:53Alert sent to the shift lead
- 14:08:54Certified system stays primary
Step 1 of 5
The image of the aisle
The camera needs to see the area from floor to ceiling, including the space above the racks, where smoke gathers first. A usable frame has enough contrast and steady exposure, with no flickering lights shining straight into the lens. Anything hidden by racks or walls stays outside the analysis.
Technical options and when they fit
General model, tuned on site
We start from a model trained on smoke and flames in general and tune it with frames from your cameras, including steam, dust or lighting specific to the site. Suits a pilot on a few cameras.
Processing next to the cameras
The exported model runs on a device with a graphics card inside the building, so alerts do not depend on the internet connection. Suits warehouses and remote sites, or those with an unreliable connection.
One server for many cameras
A single local server analyses the streams of several cameras and sends alerts to the security control room. Suits sites whose cameras are already centralised in an NVR or a VMS.
What you receive at hand-over
- The smoke and flame detection model, tuned on your images and exported for the chosen hardware
- The monitored and excluded zones drawn on each camera, with the consecutive-frame confirmation rules
- Alert channels configured per shift, with the image or clip attached and, typically, a link to the VMS or the control room
- The interface where alerts are reviewed and marked as real or false, together with their log
- A measurement report on recordings from your site, including sources that look like smoke
- Documentation, team training and a procedure for retraining with new frames
No strings attached: tell us what the cameras should see and we'll tell you honestly whether it can be done.
Let's talkWhat the visual detection system can do
- 01
Recognises smoke patterns
The model looks in the image for the shapes and movement typical of smoke, from a thin wisp to a spreading cloud.
- 02
Recognises visible flames
It flags the areas of the frame where flames appear, separately from the smoke signal.
- 03
Follows how it develops
It compares several consecutive frames to tell a phenomenon that persists or grows from a passing one.
- 04
Ignores known areas
Areas where smoke or steam is normal, such as exhaust stacks or vents, can be excluded.
- 05
Sends an alert with an image
The alert goes straight to the operator or the person in charge, with the frame or short clip that triggered it.
- 06
Keeps an alert log
Each alert and the decision the person took are recorded, for later review and for tuning the system.
How it works
From the camera image to a human check
The steps marked in yellow are approved by a person.
- Input
- AI
- Action
- Result
- Human in the loop
- 01
Input
Camera feeds
Cameras facing the sensitive areas send their images for analysis.
- 02
AI
Smoke and flame detection
The model looks for smoke and flame patterns in each frame.
- 03
Action
Confirmation over time
The signal must persist for several frames and lie outside excluded areas.
- 04
Result
Alert with image
The operator or person in charge receives the alert with visual evidence.
- 05
Human in the loop
Check and decision
A person checks the situation and follows the internal emergency procedure.
What visual detection needs
- Cameras with a clear view of the areas of interest, with resolution and lighting suited to day and night
- Real images from the site, including sources that look like smoke: steam, dust, fog, exhaust fumes
- A list of the areas to watch and of those where smoke or steam is a normal part of the work process
- The internal emergency procedure and the people who should receive alerts, by shift
- Processing resources close to the cameras, ideally on site, so alerts do not depend on the internet connection
Possible integrations
- IP cameras with an RTSP stream
- ONVIF-compatible cameras
- Thermal cameras, where they exist and provide an accessible stream
- NVRs and VMS systems
- NVIDIA Jetson edge devices
- Notifications by email, SMS or messaging apps
- REST API and webhooks
- MQTT
- IP cameras with an RTSP stream
- ONVIF-compatible cameras
- Thermal cameras, where they exist and provide an accessible stream
- NVRs and VMS systems
- NVIDIA Jetson edge devices
- Notifications by email, SMS or messaging apps
- REST API and webhooks
- MQTT
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What stays with people
- Checking every alert and deciding whether to trigger the emergency procedure
- Setting the watched areas, the excluded ones and who receives alerts on each shift
- Maintaining and regularly testing the certified fire detection and alarm systems, which remain the primary system
- Reviewing false alarms and missed events, so the system can be tuned based on them
Limits and operating conditions
- The system only sees what the camera sees: smoke hidden by racking or walls, or outside the frame, cannot be detected
- Steam, dust, fog, reflections or flashing lights can look like smoke or flames and cause false alarms
- At night, in darkness or in dense fog, visual detection becomes less reliable without suitable lighting or cameras
- It is not a certified fire detection system and does not replace the equipment, permits or legal fire safety obligations
How we evaluate smoke and fire detection
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- False alarms
- Alerts that do not match real smoke or flames, counted over a period of operation and grouped by cause
- Missed events
- Tested on footage of real smoke or flames, or of smoke simulated under safe conditions, checking whether an alert was raised
- Time to alert
- The interval between the first frame in which smoke is visible to a person and the moment of the alert, measured on the same footage
- Stability across conditions
- Results compared separately for day and night, clear weather, fog or rain, indoor and outdoor areas
How we put visual detection into service, in 4 steps
- 01
We choose the areas and the procedure
We agree which areas matter, who receives alerts and how they connect to your emergency procedure.
- 02
We check the cameras
We review the angle, distance and image quality by day and by night, and propose additions where needed.
- 03
We test and reduce false alarms
We run the system on real images, mark areas with steam or dust and tune thresholds until alerts become useful.
- 04
We integrate and monitor
We connect alerts to your team's channels and regularly review false alarms and test events.
Use cases
The scenarios below are conceptual examples, designed to illustrate realistic applications. They are not implementations delivered by WEB-AZ.
- Logistics
High-bay warehouse with racking
Cameras on the main aisles flag smoke visible above the racking, and the security operator checks the image and follows the internal procedure.
Conceptual example - Waste management
Outdoor storage platform
On a platform holding combustible materials, the alert with an image reaches the shift supervisor's phone, including at night and at weekends.
Conceptual example - Agriculture
Yard with bales and machinery
A camera covering the yard flags smoke near the bale store, and the farmer receives the clip and decides whether to step in or call the emergency services.
Conceptual example
Useful guides
Before you decide, read
Frequently asked questions
What is visual smoke and fire detection?
It is a computer vision technique in which a model analyses camera images and recognises the patterns typical of smoke and flames. When it finds them, it sends an alert with the related image to a person who checks the situation.
Can it replace smoke detectors or the fire alarm system?
No. Visual detection is an additional warning layer and is not a certified fire detection and alarm system. The equipment, permits and inspections required by fire safety regulations, including those overseen by ISU (the Romanian emergency situations inspectorate), remain mandatory and separate.
Where does visual detection make sense?
In areas that conventional detectors struggle to reach: very high spaces, yards, outdoor platforms or large warehouses. There, a camera that sees the area can spot smoke before it reaches a ceiling detector or is noticed by someone.
What about steam, dust or fog?
They can look like smoke and cause false alarms. That is why we test the system on images from your site, exclude areas where steam is normal and require the signal to persist for several frames before an alert. Any remaining false alarms are reviewed and used for tuning.
Does it work with the cameras I already have?
It can, if the image covers the area of interest and is clear enough, including at night. We check the angle, resolution and stream of the cameras in practice before confirming, and propose additions where they fall short.
Who receives the alert and what do they need to do?
The alert goes to the people you designate for each shift, for example the security operator or the shift supervisor, together with the image that triggered it. What happens next is decided by a person, following your internal emergency procedure.
Can’t find the answer? Write to us and we’ll reply directly.
Do you have areas that conventional detectors struggle to cover?
Send us a few images from the cameras that see those areas and tell us who should receive the alerts. We will tell you whether visual detection makes sense there and what conditions need to be met.
Related solutions
- Video event detectionVisual rules that flag entries into restricted zones, missing protective equipment or fallen objects, for human review.
- Operational monitoringVisual tracking of process stages and of vehicle and pallet movement in yards and warehouses, so you can see where the flow gets stuck.
- Edge AI for computer visionVideo analysis on devices installed on site: fast response, less network load and images that never leave the building.