Offline Face Recognition on ATLAS BOX

The problem

Some sites need face recognition at doors, gates and reception but cannot send face data to an external service. Regulated environments have strict rules about where personal data is processed and stored. Closed and isolated networks have no outbound route to the internet at all. Remote sites cannot depend on a connection that drops. A cloud-hosted service is ruled out in each case, while building and staffing a server room for a single entrance is out of proportion to the job.

The solution

Face recognition runs on the ATLAS BOX appliance itself, at the site where the cameras are.

• Enrols and matches faces on the appliance, so face data stays inside the customer’s own network.
• Runs fully offline: the appliance can be deployed on an air-gapped network with no outbound route to the internet.
• Handles inference on a 6+ TOPS NPU, while an ARM A76 CPU runs the applications alongside it.
• Carries video analytics and video management on the same box, so one appliance covers recognition, detection and recording.
• Holds recorded video on 4TB of local NVMe storage.
• Keeps working through a connectivity outage, because nothing in the recognition path depends on the link.
• Grows by adding appliances site by site rather than by widening a connection to a central service.

What we brought

• On-device face recognition
• Offline operation
• Edge inference
• Local video storage

How we solved it

1. 1Enrol
2. 2Capture
3. 3Match
4. 4Decide
5. 5Record