Late-night access patterns flagged at the staff door. Lanyard validation in the MICE space.

Vendor type

Boutique hotel chain (under-200-keys cohort)

Site covered

1 property — Manchester city centre

Rooms

~200 rooms

Camera estate

~52 cameras (front of house + MICE + staff bank)

Duty model

1 overnight duty manager + 1 night porter

ICO posture

ZA-class controller; embedding-only ReID; no facial-recognition store

Outcome headline (template)

Late-night staff-door pattern flagged inside the moment; MICE lanyard mismatch surfaced at the lift.

Embedding-only ReID across front-of-house, MICE space and the staff-door bank

Cameras at the front entrance, the lift bank, the MICE conference floors and the staff-door corridor all stream to the edge node. Each detected person is reduced to a vector; cross-camera reach across the three zones surfaces as a thumbnail strip in the duty manager's pane. No face image is retained; no name is computed.

Late-night access pattern flag at the staff door + lift bank

When an embedding vector enters the staff-door corridor between 23:00 and 05:00 and is then seen on a guest-floor lift, the system flags the cross-zone movement. The duty manager sees a single row — 'staff door 23:42, guest lift 23:44, no badge swipe in between' — and can act inside the moment rather than from a morning-after access-log review.

MICE-space lanyard validation

Delegates entering the MICE conference floors are credentialed by a coloured lanyard. The system pairs the embedding vector with the lanyard colour as seen on entry; when the same vector reappears on a different conference floor with a different lanyard colour visible, the mismatch surfaces at the lift. The duty manager triages — most are innocent (a delegate borrowed a friend's lanyard); some are not.

Guest-facing posture: no face image, no booking-name link

The system never crosses into the PMS. The vector is not tied to a guest name, a room number, or a booking record. Lobby signage names what the system does ('cross-camera matching by clothing and gait') and what it does not do. The guest experience is unchanged.

The named property + the GM sign-off

We will publish the property name (with operator consent), the GM's name and title, and the operator's existing CCTV DPIA marked-up against our embedding-only delta.

The measured staff-door pattern — count, false-positive rate

We will publish the count of late-night staff-door cross-zone events the system flagged in the first 90 days, the false-positive rate (the duty manager's judgement of 'flagged but routine'), and the count of confirmed incidents.

The MICE-space lanyard mismatch outcome

We will publish the count of lanyard mismatches surfaced in the first three MICE events, the triage outcome (innocent / coached / removed from the floor), and the conference-organiser feedback on whether the posture made the event easier or harder to run.

The honest miss-list

We will publish what the system missed in the first 90 days that the duty manager did catch by eye. We will not publish a success-only template.