PLOMID Cabin
In-cabin occupant intelligence, at the edge.
Camera observations become seat and occupant states, states become events with provenance, and PLOMID stores both — on the vehicle's network, offline, without a second database.
- 00:31 DRV NOT_VISIBLE → OCCUPIED
- 00:36 PAS OCCUPIED → OCCLUDED
Where Cabin is today.
Cabin is an engineering demonstrator and pilot platform, not a production-certified occupant-classification or occupant-protection system. It runs on pretrained models for inference only — no training and no fine-tuning — and any state it cannot see is reported as unknown rather than as a negative.
- Runs
- Beside a cabin camera on edge hardware, offline, and recovers from camera, model, connection or process failure.
- Stores
- State, events, configuration and diagnostics in PLOMID, through its documented SQL interface only.
- Shows
- An engineering console: live and replay operations, the seat map and its editor, the event timeline, diagnostics, validation runs, configuration, pilot dossier, fleet rollups and an on-demand benchmark card.
- Does not
- Claim OEM validation, accuracy figures, or functional-safety and privacy certification. Those come from an OEM-specific programme, not from this site.
A camera, a state engine, one database.
Capture, perception and temporal logic run on the edge. Everything durable — the current state, every transition, the configuration and the diagnostics — lands in PLOMID, where an application queries it with SQL like any other table.
- 01 Capture Frames with timestamps, bounded buffers, reconnect on loss.
- 02 Perception Pretrained detection and tracking; posture where landmarks are visible.
- 03 Seat association Tracks bound to configurable seat regions, with the association’s own uncertainty.
- 04 Temporal state Dwell, smoothing and hysteresis, so a state changes on evidence, not on one frame.
- 05 PLOMID State, events and diagnostics persisted over SQL as the system of record.
- State Current occupant and cabin state per seat region, with the uncertainty vocabulary preserved.
- Events State transitions with timestamps and provenance, so a change can be explained after the fact.
- History State and events queryable in SQL beside the rest of the application data.
- Diagnostics Model, camera and connection health, and the deliverable that lets an engineer reproduce a session.
Missing evidence is a state, not a guess.
A camera sees what it sees. Cabin keeps that honesty in its data model: an unreadable seat stays unreadable, and a downstream system can decide for itself what to do about it.
- OCCUPIED
- A seat region holds a visible occupant with a stable track.
- UNKNOWN
- Evidence is inconclusive. The state is unknown, not empty.
- OCCLUDED
- Something is in the way. A blocked view is not an absent person.
- NOT_VISIBLE
- The seat region is out of the camera’s field of view.
- UNSUPPORTED
- The requested label is outside what the models can report.
- STALE
- The last observation is too old to stand as current.
Every number on screen traces to a row.
The engineering console is the surface the program is operated through — nine destinations, from live operations to the benchmark card, and a guided tour that walks them in about a minute. Every value it displays is either a PLOMID row or a live reading, labelled as such.
One screen: stream, seat map, alerts and evidence. Every number traces to a PLOMID row or the live hub.
- near-adult-left OCCUPIED
- mid-adult-center LIKELY_OCCUPIED
- child-seat-right UNKNOWN
- Live / Replay The operations screen: stream, seat map, alerts, unique-seat counts and evidence, with session setup for source, geometry, threshold, mask quality and the optional pose and face previews.
- Seat map The cabin as a schematic for operations and as the true calibrated zones — plus the only seat editor: trace zones on a captured frame, clear the overlap warnings, save, and a running session adopts the layout in place.
- Timeline Every seat and alert event on one evidence clock. Click an event and the console jumps to the frame behind it.
- Diagnostics Throughput, latency, model, camera and database health for the running session, in one place.
- Validation Recorded evaluation runs with build, dataset and hardware. Unsupported attributes stay visible, and thresholds are never tuned on the final split.
- Config The geometry, thresholds and model provenance the console is actually running with.
- Pilot dossier A capability and gap statement generated live from PLOMID rows, downloadable as JSON or Markdown — gaps stated plainly.
- Fleet Aggregate rollups across recorded sessions. No raw media ever leaves a vehicle.
- Benchmark card Measured on demand on the machine doing the work — never pledged, never copied.
- Same layer
- The console reads the same PLOMID the pipeline writes: no sidecar store, no dashboard copy of the state.
- Replay is re-reading
- Reopen a session and seats, counts and evidence come back from the rows the live run wrote — not from a cache.
- Empty states, kept
- When a panel has nothing, it says what it has not read — “No frame yet — press Start.”, “No seat states yet.” — instead of inventing data.
Pilot it against your environment.
A pilot starts with your cabin geometry, your camera and the questions your team already asks of the data. We bring the pipeline and the database; you bring the environment it has to work in.