Video Surveillance Analytics

A camera samples a frame on a gentle cadence rather than streaming. If the scene is unchanged, the change gate reuses the last reading and spends nothing. If it changed, the vision stage reads the objects, the deterministic stages filter sub-threshold noise and NMS-dedup the boxes, track the objects across frames, and test each track's footprint against the camera's restricted-zone polygons and their class allow-lists.

Only an object that dwells inside a zone across N consecutive frames is confirmed; only a confirmed, governed situation becomes an alert - routed to the on-watch operator, who acknowledges, dispatches and resolves it. Later frames confirm the resolution, or the alert escalates up the chain on timeout.

Most of the journey is the system staying quiet, and that is the point. A shadow is dropped by the confidence floor; an animal is suppressed by a zone's class allow-list; an object whose read sits between the detection floor and the auto-alert floor is routed to manual review rather than auto-fired. A degraded or occluded frame the OSS model can't read confidently stays unconfirmed and routes to a human too - never a fabricated certainty. Fewer false alarms is the product.