Edge AI: What It Is and How It Works
What is edge AI
Edge AI means running artificial intelligence models close to where the data is generated: in a camera, in a compact device next to the machine or on a server inside the site itself. The model analyzes the image, sound or signal right there and only sends out the result, not the raw data.
The "edge" is the outer end of the network, where the sensors are, as opposed to the data center or the cloud.
How it works
An edge AI system usually has three parts:
- Sensors, such as IP cameras, scales or machine sensors, connected over the local network.
- An edge node, a device with computing power, often with a GPU or an accelerator, that runs the model. To fit and run fast, the model is optimized for that specific hardware.
- A central platform, which receives the results, stores them and shows them on dashboards.
The node turns what it sees into small events: a person has come in, a tray is empty, a part has a defect. Those events take up very little space compared with the video they come from.
Edge vs. cloud AI
The difference between edge and cloud AI is where the computing happens. In the cloud, data travels to a data center and is processed there. At the edge, it is processed on site.
The edge makes sense when:
- Latency matters and the response has to be immediate.
- The volume is high, such as several video streams, and sending it all would saturate the connection.
- Privacy calls for it and images must not leave the premises.
- The connection is not guaranteed and the system has to keep measuring without a network.
The cloud is still useful for consolidating data from several sites, training models and managing users. That is why many architectures are hybrid: the edge captures and processes, and the cloud combines and analyzes.
A simple example
A store wants to know how many people walk through each area. Instead of sending video from its cameras to the cloud, a device inside the store detects and counts people. Only the counts per time slot reach the platform.
Edge AI in DPI projects
In MBO Smart Buffet, the system DPI built for MBO, live detection and tracking run at 15 fps per camera inside the hotel, on NVIDIA Jetson devices. The video never leaves the hotel: only small events reach the cloud. The node's queue holds 20,000 events or 7 days without a connection, so an outage only delays the data. The split of work is clear: the cloud decides, the edge observes.
DPI's computer vision for retail solution follows the same principle. It processes images locally on an NVIDIA Jetson edge device, to guarantee privacy and avoid depending on the cloud.