Computer vision · Edge AI
Analysis next to the camera.Images stay here.
Edge AI for computer vision means the models run on devices installed on site, next to the cameras, instead of video streams being sent to the cloud. The response comes faster, the network carries less load, and only events and statistics leave the building. Local hardware has its own limits and needs planned maintenance, which we discuss from the start.
Sending all video to the cloud is not always the right choice
A high-quality video stream uses a lot of bandwidth. When there are many cameras, the internet is weak or the reaction has to be immediate, moving the images becomes the problem itself. Add to that a natural question: why should images leave the building if only an event matters?
- The site's connection cannot carry the streams from all cameras at once
- The round trip to the cloud adds too much delay for decisions on the line
- Internal policy or the risk assessment requires images to stay on site
- Cloud transfer and storage costs grow with the number of cameras
Who it is for
- Sites with limited or unstable internet, such as factories, farms, construction sites or remote warehouses
- Applications that need a fast reaction, for example on a production line
- Organisations that prefer images not to leave the site, for privacy reasons
- Projects with many cameras, where sending every stream to the cloud would be costly
- IT teams that need to manage devices across several sites
What an edge AI video system does
- 01
Local analysis
The model runs on the on-site device and analyses frames as soon as it receives them from the camera.
- 02
Sends only the results
Events, totals and, where needed, a relevant frame leave the site, not the full video stream.
- 03
Keeps going when the network drops
If the internet goes down, the device continues its analysis and stores results locally until it reconnects.
- 04
Models optimised for the hardware
We adapt the models, for example through quantisation, so they fit within the device's memory and power limits.
- 05
Controlled remote updates
Models and software are updated remotely, in stages, without a site visit for every change.
- 06
Device monitoring
The temperature, load and connection of each device are monitored, with an alert when something is not working.
How it works
What happens on site and what leaves it
The steps marked in yellow are approved by a person.
- Input
- AI
- Action
- Human in the loop
- Result
- 01
Input
Cameras on the local network
Compatible cameras send their stream only to the edge device on the same network.
- 02
AI
On-device analysis
The optimised model processes frames locally, without sending them outside.
- 03
Action
Filtering and local storage
Only relevant events are kept, with limited retention.
- 04
Action
Results sync
Events and statistics reach the central dashboard when the network allows.
- 05
Human in the loop
Update approval
A new model is tested and approved before it is rolled out to devices.
- 06
Result
Managed local fleet
Fast analysis at each site, with central control over the devices.
What information is needed about cameras and the network
- An inventory of cameras: model, resolution, streaming protocol, codec and position on the network
- Latency and availability requirements for each use case
- Mounting space with a stable power supply, ventilation and, where needed, an industrial enclosure
- Secure network access for remote management, approved by your IT team
- The model that will run and a set of test images to measure it on the chosen hardware
Possible integrations
- NVIDIA Jetson devices
- Industrial mini PCs with an AI accelerator
- Hailo or Google Coral accelerators
- IP cameras with an RTSP stream
- ONVIF-compatible cameras
- Docker containers for software distribution
- MQTT for sending events
- AWS IoT Greengrass or Azure IoT Edge
- Grafana dashboards
- NVIDIA Jetson devices
- Industrial mini PCs with an AI accelerator
- Hailo or Google Coral accelerators
- IP cameras with an RTSP stream
- ONVIF-compatible cameras
- Docker containers for software distribution
- MQTT for sending events
- AWS IoT Greengrass or Azure IoT Edge
- Grafana dashboards
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What stays with people
- Approving any new model or update before it is rolled out to devices
- Deciding what data may leave the site and in what form
- Regularly checking devices, especially in environments with dust, vibration or high temperatures
- Managing physical and remote access to devices, together with the IT team
Limits and operating conditions
- Edge devices have limited memory and computing power; large models or too many cameras on one device may need additional hardware
- Optimising a model for the hardware can reduce detection quality, and the difference must be measured
- Hardware in the field can fail or overheat and needs planned maintenance
- Not every camera can be connected directly; the stream, codec and network are checked for each device
- Managing several sites requires clear processes for updates, monitoring and security
How we evaluate an edge system: latency, quality, availability
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- Processing latency
- The time from receiving the frame to the result, measured on the device under real load conditions
- Quality after optimisation
- Results of the optimised model compared with those of the original model, on the same test set
- Device availability
- Uptime and outages, reported by the device monitoring system
- Volume of data sent
- Network traffic generated by each site, compared with the volume of raw video streams
How we implement edge AI at your site
- 01
We assess the site and the network
We inventory the cameras, check the streams and agree what must be processed locally and what can leave.
- 02
We choose and test the hardware
We run the model on candidate devices and measure latency, power consumption and temperature.
- 03
We install and secure
We mount the devices, configure secure access and set local data retention.
- 04
We manage the devices
We monitor them and roll out approved updates, with the option to roll back to the previous version.
Use cases
The scenarios below are conceptual examples, designed to illustrate realistic applications. They are not implementations delivered by WEB-AZ.
- Livestock farming
Analysis on a farm with weak internet
A device on the farm analyses the cameras in the animal housing and sends only alerts and daily statistics, without loading the mobile connection.
Conceptual example - Manufacturing
Fast reaction next to the line
The industrial computer next to the line processes images locally, so the alert reaches the operator before the part leaves the station.
Conceptual example - Multi-site retail
The same solution in several stores
Each store has its own device, which sends only anonymous totals, and the central team updates the model in stages, store by store.
Conceptual example
Useful guides
Before you decide, read
Frequently asked questions
What is edge AI in computer vision?
Edge AI means running video analysis models on a device located on site, close to the cameras, rather than on cloud servers. It is an architecture choice, not a separate use case: visual inspection, counting or video alerts, for example, can all run on the edge device.
Does edge AI mean images never leave the site?
It can, if it is configured that way. We agree together what leaves the site: usually only events and statistics, and possibly a relevant frame. Everything else is processed and deleted locally, in line with the agreed retention.
What happens if the model on the device is not sure of a result?
Low-confidence results can be marked and sent, with the related frame, for human review. If such cases become frequent, it is a sign that the model or the filming conditions need to be reviewed.
What hardware do you recommend?
It depends on the number of cameras, the complexity of the model and the environment. We test on candidate hardware before recommending anything, because specifications on paper do not tell the whole story about real performance.
Who handles maintenance?
We set it out in the contract: monitoring, updates, replacing faulty equipment and which interventions are carried out on site. An edge system left without maintenance degrades over time.
Is it better value than cloud processing?
Not always. Edge reduces cloud transfer and storage, but adds hardware and maintenance. We compare both options, or a combination, for your specific case.
Can’t find the answer? Write to us and we’ll reply directly.
Want your video analysis to run on your own site?
Tell us how many cameras you have, what your network looks like and what must stay local. We assess whether edge, cloud or a combination makes sense.
Related solutions
- Industrial visual inspectionCameras on the production line flag parts with visible defects, and the operator confirms every rejection.
- Video event detectionVisual rules that flag entries into restricted zones, missing protective equipment or fallen objects, for human review.
- AI evaluation and monitoringTest sets, checks at every change, human spot-check reviews and cost tracking, so you know how your AI performs over time.