Building SentinelVision: threat detection that refuses to recognize faces

SentinelVision is our multi-camera detection system, in active development: it watches video feeds for weapons, people, and vehicles, and raises tiered alerts when something needs attention. It runs on Raspberry Pi hardware, on-site, and it deliberately cannot tell you who anyone is.

The design constraint that shaped everything

We decided early that facial recognition was out. Not "not yet," out. This isn't a limitation we apologize for; it's the spec.

The reasoning is simple: the question a security system needs to answer is "is there a weapon in frame?", not "which student is this?" The first question protects people. The second builds a database of identities, movements and false matches that becomes a liability the moment it exists. A system that never extracts identity has nothing to leak, nothing to subpoena, and nothing to quietly repurpose later. The strongest privacy guarantee is data that was never collected.

Making it work on a $80 computer

The harder engineering problem is the hardware. Cloud vision APIs are easy; running detection on a Raspberry Pi at the edge, with no cloud and no video leaving the building, is not.

• Two models, two jobs. A fast, lightweight pass runs continuously on every frame looking for people and motion. The heavier weapon-detection model only spends its compute where the fast pass found something worth looking at. Watching an empty hallway is cheap; scrutiny is reserved for frames that earn it.
• YOLO-family models compiled through NCNN, an inference framework built for exactly this class of ARM hardware. Quantized to run on CPU on the Pi.
• Tiered alerts. A person after hours is a note. A possible weapon is a different tier entirely, pushed immediately. The worst failure mode for a detection system isn't a missed frame. It's the operator who learned to ignore it because everything was an alarm.

Edge-first is privacy-first

Nothing streams to a cloud service. The dashboard is only reachable over a private network, not "protected by a login page on the open internet," but not on the open internet at all. Footage stays in the building; what leaves is an alert.

SentinelVision is in active development, and the product page has current status and a beta notification signup if you want to know when it's ready for real deployments.