Edge AI or cloud: where should your video streams be processed?

Definitions and typical architectures

Every video analytics solution has to answer an architectural question: where is the footage analysed? The answer affects the network, security, data protection and day-to-day operations.

There are three main architectures:

• Edge AI (on-site processing): a compact computer fitted with an AI accelerator is installed on site, close to the cameras. It reads the video streams locally, runs the detection models and only sends out results: counts, events, durations, alerts.
• Cloud: video streams, or clips, are sent to remote servers where the analysis takes place. The site only hosts the cameras and the network.
• Hybrid: analysis runs on site, and aggregated results are consolidated in a central platform that provides multi-site dashboards.

Bandwidth and site connectivity

A video stream good enough for analysis is a continuous data flow whose size depends on resolution, frame rate and compression. Multiply it by the number of cameras and it can quickly exceed the upload capacity of a site’s internet connection, especially for shops, branches or service stations outside major cities.

Edge processing removes that problem: streams only travel over the local network, between the cameras and the analytics unit, and just a few kilobytes of results go upstream. It also keeps working if the connection drops: results are buffered and sent once the link is back.

The cloud, on the other hand, requires an upload connection that is sufficient, stable and sized for the long term. It can work for a handful of cameras on a well-connected site, or for occasional analysis of clips, but becomes harder to sustain at scale.

On the edge side, sizing the unit depends on the number of streams to analyse, the resolution and analysis frame rate chosen, and the models being run. It is worked out during the feasibility study, based on the site’s actual cameras rather than a catalogue: an undersized unit drops frames, while an oversized one is money wasted.

Data protection and where footage lives

Images of identifiable people are personal data, so where they are processed and stored is not a mere technical detail. With edge processing, footage stays on the customer’s premises: it never leaves the site, and only aggregated indicators, with no images or identification, are transmitted.

With cloud processing, transfer and storage need careful framing: stream encryption, server location, provider access, retention periods. Hosting outside Morocco may trigger specific formalities. In every case, the company operating the cameras remains the data controller under Law 09-08 and must complete any applicable formalities with the CNDP.

Edge processing does not remove these obligations, but it reduces how exposed the footage is and makes proportionality easier to demonstrate.

Maintenance, updates and fleet monitoring

The argument most often made for the cloud is easier maintenance: one environment to update, no analytics hardware to manage on site. It is a fair point, and one that an edge architecture has to take seriously.

A fleet of edge units must be managed like any IT estate. Things to plan for from day one:

• remote monitoring: the status of each unit, temperature, load, lost or frozen camera streams;
• remote rollout of software and model updates, with the ability to roll back to the previous version;
• security: encrypted remote access, unique passwords, an isolated video network, patches applied;
• installation conditions: a dry, ventilated room, protected power and ideally a UPS;
• a fast replacement procedure in case of hardware failure.

Hybrid architecture: local processing, central consolidation

In practice, the question is not always edge or cloud, but what runs where. A hybrid architecture combines the strengths of both: footage is analysed on site, and aggregated indicators are consolidated in a central platform that lets you compare sites, keep history and share dashboards.

That is the approach VIDEO.MA favours: edge processing wherever possible, with results consolidated for management reporting. Today, projects are delivered case by case after a feasibility study; a multi-site SaaS platform is being prepared for a later phase and is not yet available.

In a hybrid set-up, you also need to decide up front whether any images may be sent upstream on occasion, such as a snapshot to verify an alert or recalibrate a zone. These exceptions should be defined, limited, documented and consistent with the declared purposes; by default, only aggregated results leave the site.

Making the call for your sites

A few questions help point you in the right direction:

• how many cameras per site, and what upload bandwidth is actually available?
• do your sites need to keep working during an internet outage?
• can footage leave the premises, given your commitments and formalities?
• is there a suitable technical room for an edge unit?
• who will monitor and maintain the fleet over time?

The bottom line

For most physical sites, edge processing strikes a good balance between network load, resilience and data protection, provided monitoring is properly organised. The cloud still has its place for consolidation and dashboard access.

The choice is not set in stone either. A network can start with a few sites processed at the edge, then add central consolidation once the indicators have proved useful. What matters is deciding early where footage is processed, what leaves the site and who maintains the equipment, because those choices shape the network, the data protection formalities and the operating costs for years to come.

To dig deeper, visit our Technology and Hardware pages, or ask us to review your architecture: we will look at your sites, network and cameras before recommending an approach.