Edge Analytics or Server-Based Analytics? Choosing the Architecture

Edge video analytics, on-camera and GPU server analytics: a comparison in terms of model flexibility, network load, maintenance and licensing, a decision tree and a hybrid example.
Technical summary
Technology
IP camera, edge AI, GPU server, RTSP, ONVIF Profile M, VMS
Topic
Choosing a video analytics architecture
Level
Intermediate–advanced
Prerequisites
  • Familiarity with IP camera and VMS architecture
  • Basic network and bandwidth knowledge
In this article9 sections

In analytics projects, the first discussion often starts with the wrong question: “Which software is better?” The real question is where the model will run: on the camera’s own processor, on an edge box at the site, or on a GPU server in the data center. This decision determines model flexibility, network load, maintenance burden, licensing cost and how the system will grow, all at once; and it’s expensive to change later. Below we compare the three layers against the same criteria and give a decision tree and a hybrid architecture example.

Layers where analytics can run

Every analytics workflow passes through four steps: decoding the image or taking sensor data directly, preprocessing (scaling, color conversion), running the model (inference), and sending the result to the rules engine, VMS or integration layer. Where each of these steps happens defines the architecture.

  • On-camera: The model runs on the NPU/DSP in the camera’s SoC. What leaves the camera isn’t video but events and metadata.
  • Edge box (intermediate layer): A small server or embedded device with a GPU or accelerator card that processes the streams of a site’s cameras locally. Channel capacity varies by model; typically tens of channels.
  • Central server: Analytics running on GPUs in the data center or main campus, on the same servers as the VMS or separate ones.

These layers aren’t mutually exclusive. The right question on site is: which analytics task runs in which layer with the least risk and the lowest total cost? In most enterprise projects, the answer isn’t a single layer.

On-camera analytics

Most current mid-range and high-end IP cameras run line-crossing, area intrusion and loitering rules with person/vehicle classification on the camera itself. Because analysis is done on the frame before compression, it isn’t affected by codec degradation; that’s an advantage with distant, small targets. But don’t overstate it: the camera processor often processes the frame after shrinking it to the model’s input size. An 8 MP sensor doesn’t mean 8 MP analysis.

Strengths: No extra hardware, only events and metadata go to the network, and processing capacity grows automatically as cameras are added. When the site connection drops, the camera can keep generating events and recording to its SD card.

Limits: Model options are limited to what the manufacturer offers; training a special object class, such as a particular type of equipment, is often not possible. Because processing power is constrained, the number of rules running at once and the analysis frame rate are also limited. On mixed-brand sites, the real problem is the event format: each manufacturer sends its own structure, and these have to be reduced to a single standard on the VMS side. ONVIF Profile M defines a common framework for analytics metadata and events; but the level of support varies from device to device, and “Profile M” on a datasheet isn’t enough on its own. You can find the differences between profiles in detail in the ONVIF Profile T, G and M comparison.

Edge box and intermediate layer

An edge box comes in at sites where you want to add analytics without replacing cameras, or where on-camera models aren’t enough. The box usually pulls the substream or main stream from cameras via RTSP, decodes the frames and runs its model.

Strengths: It’s independent of camera brand; older cameras are brought into analytics too. Model updates are done on a single device. Because video is processed without leaving the site, only events, clips and metadata go to the WAN; in multi-site structures, it preserves WAN bandwidth.

Limits: Each site gets one more server needing maintenance, updates, spare parts and physical security. Because it works on a compressed stream, low bitrate, night noise and aggressive compression lower detection quality. A picture we often see on site: the box runs at the bottom of a rack in an airless cabinet, the GPU heats up, the frame rate drops and nobody notices. If the box stops, all of the site’s analytics stops. Monitoring and a backup plan must be designed from the start.

Central server analytics

In a central architecture, all analytics runs on GPU servers in the data center. It’s the natural choice for large campuses and setups where cameras already record to a central VMS.

Strengths: Model flexibility is highest here: heavy work such as different models, custom-trained classes, attribute-based search and multi-camera tracking is done with central GPU resources. Capacity is managed in a single pool, redundancy is built with N+1 servers, and version control is from a single point.

Limits: Every analyzed stream is carried over the network to the server. If the recording stream already comes to the center, the extra load is limited; but if a separate stream is pulled for analysis, it must be added to the uplink and backbone calculation. This is the item most often forgotten in projects. Do the calculation with the method in the camera network bandwidth calculation guide. At remote sites, WAN latency and outages directly affect analytics: when the line goes down, so do the alarms. We covered server capacity by stream, FPS and model variables in the GPU server sizing guide.

Comparison table

Criterion On-camera Edge box Central server
Model flexibility Low, depends on the manufacturer Medium–high High
Network load Very low (events, metadata) High on the LAN, low on the WAN As much as the analysis stream
Scaling Linear with cameras Per site With a GPU pool
Maintenance effort Firmware management A server per site Central, single point
Licensing model Usually included in the camera price Per channel or per device Per channel, GPU or module
Connection outage Rule runs; event delivery waits Continues within the site Stops at the remote site
Image quality Frame before compression Compressed stream Compressed stream

The row that varies most from project to project is the licensing model. When comparing quotations, add up per-channel licenses, per-GPU licenses, extra fees per analytics module and annual maintenance/subscription fees over the same period (for example 5 years). On-camera analytics looks “free”; if the cameras have to be replaced, the cost has simply moved to the hardware line.

Decision tree

The sequence below gives a quick first decision in most enterprise projects. Each “yes” brings the relevant layer forward. Ending up with more than one layer is normal; it points to a hybrid architecture.

  1. Are the required analytics covered by the manufacturer’s standard rules (line crossing, area intrusion, loitering, person/vehicle class)? If yes, and cameras are being bought new, the first candidate is on-camera analytics.
  2. Is a special object class, attribute-based search or multi-camera correlation needed? If yes, a central server or a powerful edge box is needed.
  3. Is connectivity between sites limited or intermittent? If yes, analytics should stay on site: on-camera or an edge box.
  4. Will existing cameras stay, and are they mixed brands? If yes, an edge box or central server gives more consistent results than on-camera analytics.
  5. Is alarm latency critical (such as an instant audible/visual warning on site)? If yes, the decision should be made on site. Choose on-camera or an edge box instead of a chain that goes to the center and back.
  6. Is it possible to host and maintain a server on site? If not, consider a combination of on-camera analytics and central processing instead of an edge box.

Hybrid architecture example

Take a manufacturing company as an example: a main campus and three small warehouses. The main campus has 200 cameras, and each warehouse 20–30. Requirements: a virtual fence alarm at night for perimeter security, person detection in the warehouses and attribute-based search after incidents.

  • Perimeter cameras: On-camera line-crossing and area intrusion rules run on newly purchased outdoor cameras. Events reach the center via ONVIF or the manufacturer’s VMS integration; during a WAN outage, the rules keep running.
  • Warehouses: Because the existing mixed-brand cameras won’t be replaced, an edge box is placed in each warehouse. The box does person detection on substreams and sends only events and short clips to the WAN.
  • Main campus: A central GPU server generates attribute-based metadata from the recording streams; the incident review team searches by clothing color or vehicle type.

In this design, the hard part isn’t the hardware but the event model. If events from three layers aren’t reduced to a single schema, the operator has to learn three different alarm languages. A common schema like the one below hides the difference in source in the VMS and integration layer:

{
  "source_layer": "edge",
  "site": "warehouse-2",
  "camera_id": "W2-CAM-07",
  "rule": "person-detection-dock",
  "object_class": "person",
  "confidence": 0.81,
  "timestamp": "2026-03-14T02:17:45+03:00",
  "clip_ref": "vms://W2-CAM-07/20260314T021740/15s"
}

The field names are illustrative; the source_layer field takes one of the values camera, edge or server. Adapt the schema to the event format of the VMS or integration platform you use. Two points mustn’t be skipped: the timestamp is kept with the time zone, and all layers are connected to the same time source via NTP. If the clocks drift, incident review collapses.

Common mistakes

  • Choosing analytics according to the camera list: The scenario and expected accuracy are defined first; the layer is chosen after that. The reverse means paying for licenses on the wrong hardware.
  • Not checking substream resolution: Edge box and server analytics often work with the substream; on a 640×360 substream, distant targets shrink to a few pixels. The first thing to check in the console is which stream the analytics is attached to.
  • Leaving the event format for later: If how events from different layers appear in the VMS isn’t settled in the pilot, an inconsistent alarm structure emerges on the operator’s screen.
  • Not monitoring the edge box: Whether the box is up, GPU temperature and stream connections should be connected to the central monitoring system. An unmonitored box is a failed box.
  • Mistaking false alarms for a layer problem: Most false alarms come from rule and scene design; for tuning methods, see the guide to reducing false alarms in video analytics.

Verify the architecture in a pilot, not on paper: choose a few representative cameras from each layer and compare the same scenes for two to four weeks, including nights and rainy days. For deployment and pilot scope, see our AI video analytics deployment and person, vehicle and object detection pages. The primary source for the ONVIF analytics profile is the ONVIF Profile M page.

Frequently asked questions

Is on-camera analytics less accurate than central server analytics?

Not always. Because on-camera analytics works on the uncompressed frame, it can have an advantage with small targets; but processing power and model variety are limited. Accuracy is determined more by camera geometry, scene conditions and rule design than by the layer.

How much does WAN bandwidth fall when using an edge box?

Video isn’t carried over the WAN for analytics; only events, metadata and, when needed, short clips go to the WAN. If recording is also kept on site, WAN traffic is largely limited to live viewing and event clips. Your live viewing and archive copy policy determines the exact value.

Can central analytics be done with existing cameras?

If the cameras can provide standard streams via RTSP or ONVIF, most central analytics platforms process them. What matters is the resolution and frame rate of the analyzed stream and the pixel size targets occupy in the frame. Verify this in a pilot.

Can alarms be gathered on a single screen in a hybrid architecture?

Yes, but each layer’s events must be converted into a common event model in the VMS or integration layer. The camera, rule, object class, confidence score and timestamp fields must be consistent, and all devices must be connected to the same NTP source.

Support

Contact us for support or a discovery session on this topic.

Contact form
[email protected]

Eretek Video Surveillance

IP cameras, recording and central management infrastructure. We plan blind spots, retention periods and the camera network around the risks of each site.

Knowledge Center

Other articles