EdgeAI Facial Recognition System

How it works

Recognition happens on an NVIDIA Jetson unit installed at the site, not on a server somewhere else. OpenCV takes frames from the camera and finds faces. A TensorFlow Lite model turns each face into an embedding, a short list of numbers, and that is compared against the embeddings of enrolled people stored on the same device.

Because the comparison is local, video never leaves the building. Only the result, who was recognized at which door and when, is sent to the access control system. That design is where both the speed and the privacy position come from: there is no network round trip to wait for, and no footage in anyone's cloud.

The hard part

Accuracy in a lab and accuracy in a lobby are different things. Backlighting from glass doors, people walking instead of standing still, and partial occlusion all push error rates up. Camera placement and enrollment quality turned out to matter as much as the model, so the enrollment flow rejects poor reference images instead of accepting whatever it is given.

Running on edge hardware also means the model has to fit a fixed compute budget. It was quantized and benchmarked on the Jetson itself, since figures from a development workstation say little about behavior on the target device.

This project is part of our AI Agent Development work.