The short answer
Put detection where the video is when links are long, shared or unreliable, when alerts must arrive in seconds, or when evidence must survive an outage. Put analytics on central servers when every camera has dedicated fibre, the analysis needs many cameras at once, and the team can run a data centre. In a tender, write the outcomes you need and let bidders show how their design meets them.
The two designs
An ITMS in a city or an ATMS on a highway has the same chain: cameras capture video, AI models detect vehicles, plates, violations and incidents, operators review the results, and approved violations go to the e-challan system. The architecture question is where the AI models run.
In an edge design, the models run on the camera's own processor or on a processing unit in the roadside cabinet. The site sends events, evidence images and short clips to the command centre. Live video is available on demand.
In a server design, cameras send continuous video over the network to servers in a data centre or command centre. The servers decode the video and run the models for every camera.
For the engineering detail, including latency and failure modes, read edge vs cloud ANPR architecture on our blog. This guide covers the procurement side.
Comparison for buyers
| Criterion | Edge processing | Server processing |
|---|---|---|
| Network bandwidth per camera | Low: events and evidence, video on demand | High: continuous video to the servers |
| Behaviour during a link outage | Detection continues, evidence stored and sent later | Detection stops unless video is recorded locally and re-processed |
| Time to alert | Set by the device, independent of the link | Adds network and server queue time |
| Hardware cost | Higher per camera or per site | Lower per camera, higher in the data centre |
| Power and space at the site | Processor in the camera or cabinet | Camera only at the site, servers centrally |
| Model updates | Pushed remotely to every device | Updated once on the servers |
| Analysis across many cameras | Needs central software that combines events | Natural, because all video is in one place |
| Scaling to more cameras | Each camera brings its own compute | Server and network capacity must grow |
| Data leaving the site | Mainly events and evidence | All video |
Questions that decide the design
How good is the network at every site?
Map each site's link: dedicated fibre, shared fibre, leased line or mobile data. A corridor where the fibre is cut a few times a year by road works needs detection that survives the cut. NHAI's ATMS standards expect equipment to communicate over optical fibre, with provision for 5G later, but even fibre corridors have outages.
How fast must an alert arrive?
Incident alerts for a stalled vehicle in a live lane, wrong-way driving or a crash are worth more the earlier they arrive. Write a maximum time from event to operator alert into the tender, and test it during acceptance with the link loaded.
What must happen to evidence?
Evidence for an e-challan must show the vehicle, the plate, the time and the place, and must be traceable to the device that captured it. Ask where evidence is created, how it is protected from change on its way to the command centre, and what happens to it during an outage.
Who will run the system for ten years?
A server design concentrates work in a data centre: servers, storage, cooling and software licences. An edge design spreads it across the field: devices that need remote monitoring and updates. Check which skills the operator has, and price the maintenance for the full contract.
When a hybrid design fits
Most large projects use both. Detection of plates, violations and incidents runs at the edge, so it is fast and survives outages. Central software combines the events: section speed between two gantries, searches across the corridor, reporting, operator review and links to e-challan and other government systems. Video is pulled from the edge only when an operator needs it.
Tender and procurement checklist
- State the maximum time from event to operator alert, and how it is tested at acceptance.
- State the required behaviour during a link outage: detection, local storage period, and sync after recovery.
- Ask for the bandwidth per site in normal operation and during live viewing.
- Ask how evidence is created, stamped with time and location, and protected from change.
- Ask how AI models and firmware are updated, and whether updates can be rolled back.
- Ask how device health is monitored, and what alerts the operator gets when a device fails.
- Require open interfaces: ONVIF, RTSP and documented event APIs, so the command centre is not tied to one camera vendor.
- Ask where video and evidence are stored, for how long, and who can access them.
- Price power, connectivity, data centre hardware and maintenance over the contract, not only devices.
- Ask for a reference site with a similar network and scale.
How Stellarview approaches it
Stellarview uses a hybrid design. Analytics run at the edge: on the Recon Global Shutter Camera and on processing units at the gantry for VIDS, which keeps detecting and logging incidents without internet and syncs when the link returns. EdgeTrack is the central platform for violations, incidents, device health and e-challan workflows, with cloud or on-premises deployment. AI Search runs its vision language model on the camera's edge processor.
Frequently asked questions
What is the difference between edge and server analytics in an ITMS?
In an edge design, the camera or a roadside processing unit runs the AI models and sends events and evidence to the command centre. In a server design, cameras stream video to servers in a data centre, which run the models. Many systems combine both.
Which design costs less?
It depends on the site. Edge processing costs more per camera but needs less bandwidth and less data centre hardware. Server processing costs less per camera but needs strong, continuous links and more central compute. Compare the total cost over the contract period, including connectivity and power.
What happens to evidence when the network fails?
With edge processing, violations and incidents are detected and stored locally, then sent when the link returns. With server processing, video that does not reach the server is not analysed, unless the camera records locally and the system re-processes it later.
Can an edge system still give operators live video?
Yes. Edge cameras usually provide live streams on demand through RTSP or a video management system. The difference is that continuous high-resolution video does not have to flow all the time for detection to work.
How should a tender specify the architecture?
Specify outcomes rather than a design: maximum alert time, behaviour during a link outage, evidence integrity, bandwidth per site, and how models are updated. Then let bidders propose edge, server or hybrid designs that meet those outcomes.