AI Video Analytics.

The question worth answering is not whether an analytics platform is camera-agnostic — every vendor says it is — but whether a specific estate can carry analytics at all. That begins with stream access. An ONVIF profile gives a third-party application a documented way to discover a camera and pull video from it; RTSP is the session-control protocol that negotiates and controls that stream, while the media itself travels over RTP. Two points are routinely misstated. ONVIF conformance is declared by the member on the basis of its own testing — ONVIF does not test the product — so two conformant devices do not automatically interoperate; compatibility is something we establish by testing on the actual estate, model by model and firmware by firmware. And a Profile S or Profile T stream carries video only. It does not carry the camera maker's own AI events or object metadata, which travel over that maker's SDK or event channel, so an estate bought for its on-camera intelligence may be unable to hand that intelligence to anyone else.

Beyond the protocol, the blockers are prosaic and they decide projects. Codec support — which H.264 profile is in use, whether H.265 is present and in which variant — decides whether the analytics engine can decode the stream at all. The camera's authentication method decides whether a third party can open it; ONVIF has said Profile S is being deprecated — after 31 March 2027 no new product or firmware version can be submitted for Profile S conformance, in part because username-token authentication no longer reflects current cybersecurity recommendations. Then there is capacity: a camera serves a limited number of concurrent streams, and every stream the analytics tier takes is one competing with the recorder, the operator wall and the mobile client. Resolution and frame rate have to clear the floor the rule needs. Region-locked SKUs and firmware left untouched for several years close the remaining doors. None of this is legible from a brochure — it comes out of a survey.

Where the inference runs is a design decision with real trade-offs, not a preference. On-camera inference costs no extra stream and no bandwidth, survives a link outage, and gives the lowest latency — but the analytics set is fixed by the camera's own firmware and applies only to cameras capable of hosting it, which on a retrofit estate is usually a minority. An on-premise server decouples the analytics from the camera generation and keeps video and metadata inside the building, but it is sized by concurrent decoded streams rather than camera count, and decode capacity is the ceiling people forget to check. Cloud inference removes the server but moves the constraint to per-site upload bandwidth — the figure that matters is upload, not the download speed on the broadband invoice — and it raises questions about where footage and metadata are held and for how long. In every architecture, local recording at the site must continue through a link outage and back-fill afterwards; anything that depends on the link for the recording itself is a gap dressed up as a feature.

Accuracy is governed by the scene, not by the algorithm alone. Camera angle and mounting height, pixels on target — IEC 62676-4 describes this as the DORI criteria, distinguishing detection from observation, recognition and identification by pixel density across the target — lighting and IR behaviour at night, rain, fog and dust, occlusion by shelving, vehicles or crowds, frame rate, and how tightly the rule is written all move the result more than the choice of platform does. An overview camera mounted high and wide will detect motion and will not reliably classify a person; no amount of software fixes the geometry. Any honest deployment includes a tuning period after go-live in which rules are revised against reviewed false positives, and any deployment will produce false positives — the objective is a rate the operators can live with, not zero. That is also why the operator surface matters: events belong in the VMS or command centre the security team already uses, with role-based permissions, acknowledgement and an audit trail, and with forensic search available across the estate. On multi-site work, head office watching branches is a federation question before it is an analytics question — mixed-brand estates federate unevenly, and per-site upload, local recording continuity and operator permissions govern what a central operator can actually see. We integrate that layer against the CCTV, access-control, intrusion and networking scope rather than bolting a second, parallel system onto the building.