parking_solution/wiki/entities/lpr-camera.md

---
type: entity
tags: [parking, hardware, readers, offline-first]
sources: [parking-system-architecture]
updated: 2026-07-07
---

# LPR Camera

License-plate-recognition camera. For **casual/transient** vehicles, the **plate acts as ticket +
an independent record**. (See [[parking-system-architecture]] §8, §9.)

> **Superseded direction (2026-06-15):** recognition now runs **host-side** on snapshots from
> ordinary Hikvision/Dahua cameras via the [[opencv-anpr-service]], **not** on a dedicated edge-AI
> LPR camera — see [[vision-service]]. The edge-AI camera below is kept as the original assumption /
> a fallback option, but is no longer the planned path. The host-side service also does **vehicle
> verification** (anti-plate-spoofing), which an edge-LPR camera does not.

- **Edge AI (original assumption)**: recognition runs **on-device**, so it keeps working with no
  internet — fits [[offline-first]].
- It's a **host-side** identity source: only the host sees the read; the host decides and
  commands the relay open (the [[uhppote-controller]] is demoted to a commanded relay for that
  lane). See [[entry-exit-readers]].
- Being host-in-the-loop is **good for fraud detection** — you get two independent records (the
  host's signed [[append-only-event-chain]] entry + the controller's remote-open event) that
  should reconcile one-to-one; any mismatch is an anomaly.
- Mounting: within ~15° of vehicle travel at a controlled chokepoint for best reads.

## Snapshot driver (entry/exit fraud-control record)

Separate from edge-AI LPR: the camera driver (`packages/devices/src/drivers/camera.ts`) does
**snapshot-on-event** — the host pulls a still over HTTP when an entry/exit fires and stores it,
referenced from the signed [[append-only-event-chain]] entry as an independent record. The camera
**pulls, it does not push** — so it is NOT `pushesToBackend` and the setup wizard correctly hides
the "Backend push IP" field for it (gated on the driver's `pushesToBackend` flag; only
[[dingtian-relay]] sets it).

- **Hikvision** uses **ISAPI**: `GET /ISAPI/Streaming/channels/<id>/picture` (`101` = ch1 main
  stream) with **HTTP Digest** auth. The "Enable Hikvision-CGI" toggle (Network → Advanced →
  Integration Protocol) is a *different* legacy CGI surface — **not** needed for ISAPI.
- **Dahua** uses CGI: `GET /cgi-bin/snapshot.cgi?channel=<n>` (0-based channel; the wizard's
  1-based channel is decremented).

**Driver / storage boundary:** the driver FETCHES the image bytes (client-side HTTP Digest in
`drivers/http-digest.ts`) and returns them on `Snapshot.bytes`; **storage is the caller's job**
(the future entry/exit flow stores the bytes + mints a durable `imageRef`). This keeps the device
adapter free of any filesystem/blob-store dependency. `healthCheck()` is honest — it actually pulls
a frame (exercising reachability + auth + path/channel in one shot), not a fake `ready/stub`.

### Verified on hardware (2026-06-15)

A **Hikvision** unit ("Camera 20", MAC `94:e1:ac:…`, Hikvision OUI) at `10.0.10.121`, creds
`admin` / `admin123` (Digest), TCP 80:

- Initial `curl` test confirmed the ISAPI path returns a 2688×1520 JPEG (~306 KB).
- The **real driver** (no longer a stub) was then run end to end against it:
  `healthCheck()` → `ready` (pulled a frame), `captureSnapshot()` → valid `image/jpeg`, ~322 KB,
  correct JPEG magic. Digest handshake works through `HttpCamera`.
- Reaching it from the WSL dev box required forcing the source address (`config.localAddress`,
  threaded into the driver) — see [[wsl-dev-networking]] (multi-subnet source-selection trap).

### HTTP 503 "Device Busy" — can be PERSISTENT; the real fix is stream selection (2026-06-26)

The snapshot endpoint returns **HTTP 503** with the ISAPI body `statusCode 2` / `"Device Busy"` /
`subStatusCode deviceBusy` (occasionally **500**). It comes in two flavours, and they need different
fixes — **don't assume it's a momentary blip**:

- **Transient** — the encoder is briefly occupied (another snapshot in flight, a stream starting).
  Clears on retry within a frame or two.
- **Persistent** — the **MAIN-stream encoder is saturated** and 503s on EVERY main-stream snapshot.
  Confirmed on hardware (**DS-2CD1047G3H-LIU**, 2026-06-26): `channels/101/picture` → 503 on five
  consecutive probes 800 ms apart, while **`channels/102/picture` (the SUB stream) → 200 every time**,
  a clean ~15 KB JPEG. So the path/API was correct (the camera answered with a structured Hikvision
  status); the main encoder was simply never free. A retry loop **cannot** fix this — it just delays
  the failure.

> **Sharper finding (2026-06-27, same DS-2CD1047G3H-LIU).** Re-probed `10.0.10.13` directly after a
> camera reboot with every web/live-view connection closed. **Sub (102): 10/10 rapid back-to-back
> pulls → 200** (~50 ms, ~14.7 KB) — flawless even with NO delay, harder than the live ANPR cadence.
> **Main (101): 3/3 → 503 in ~20 ms** — an *instant* reject, not a timeout. So main isn't merely
> "saturated/busy" on this model — its snapshot endpoint is **structurally unavailable**; **sub (102)
> is mandatory**, not just preferable. SEPARATELY, the camera 503s on **any** stream when its
> **connection slots are exhausted** — a human holding the web UI / live-view, or parallel
> main-stream experiments, consume slots; a **reboot clears stuck slots**. This was the actual cause
> of the **2026-06-27 "subscribers auto-enter but don't auto-exit"** scare: manual main-stream
> testing held the exit camera's slots → the system's sub-stream snapshot pulls got "Device Busy" →
> the ANPR exit read never got a frame → no auto-exit. **NOT a code/flow bug** — the subscription
> exit logic, camera→relay binding, and `stream: "2"` config were all correct (auto-exit/entry pairs
> were clean before the test storm and after the reboot). (Also recorded as the LLM memory
> `g3h-main-stream-snapshot-503`.)

**The fix that actually works: snapshot from the SUB stream.** The Hikvision ISAPI channel id is
`<channel><stream>` (e.g. ch1 main = `101`, ch1 **sub = `102`**). The driver now has a **`stream`
config field** (`1` = main, default for back-compat; `2` = sub). Set the G3H camera to **Sub (02)** in
the setup form → its status flips `degraded → ready` (verified live: pulled a 14.7 KB JPEG in ~87 ms).
The sub-stream is also the better fit for snapshot/ANPR anyway (smaller/faster; doesn't contend with
live-view/recording for the main encoder).

Two more complementary mitigations (both BUILT, for the *transient* case):
1. **Don't cause concurrent busy.** On a vehicle entry two server paths used to snapshot the same
   camera at once (the ANPR bridge + the advisory `snapshotAsync`); the 2nd concurrent GET drew a 503.
   They now share ONE pull via `captureSnapshotShared` (deviceId-keyed, `apps/server/src/snapshot.ts`)
   — the main cause of the slow 2026-06-25 subscriber entry. See [[lane-presence-and-anpr-entry]].
2. **Retry a transient one.** `HttpCamera.captureSnapshot` retries 503/500 with a short linear backoff
   (250/500/750 ms, ≤4 attempts), then fails naming it `(device busy)`; it does NOT retry 401/404
   (config errors won't self-heal). This recovers a momentary blip but, by design, still fails a
   PERSISTENTLY-busy main stream — the cue to switch that camera to the sub-stream.

Covered by `packages/devices/src/drivers/camera.test.ts` (retry behaviour + the main/sub path
selection). `healthCheck()` deliberately reports a live 503 as `degraded` (it surfaces a genuinely
saturated main stream rather than hiding it behind a retry).

## Clock sync — the 1970 power-cut reset (built 2026-07-07)

Field observation (park-buzi): after a power cut these cameras come back with their clock at the
**1970 epoch** (no/dead RTC battery, no NTP) and stay there until a human logs into the web UI —
Hikvision's web login silently pushes the browser clock. A wrong camera clock corrupts the OSD
timestamp burned into every snapshot (the evidence trail) and the times on ANPR pushes.

Built: the HOST is the site's time authority (offline-first — no NTP infra dependency), and the
device monitor re-syncs each Hikvision camera over the same Digest-auth ISAPI used for snapshots:

- **Trigger:** the camera's **offline → ready transition** (exactly the power-restored moment) +
  a **24h backstop**; the attempt timestamp is stamped BEFORE the async call, so a failing camera
  retries at backstop cadence, never every 8s poll.
- **Mechanics** (`HikvisionCamera.syncClock`): `GET /ISAPI/System/time`, parse `localTime`; drift ≤
  60s → leave alone. Beyond that → `PUT /ISAPI/System/time` with `timeMode=manual`, the site's
  wall-clock now WITH explicit utc offset (`localIsoWithOffset(siteTz)`, e.g.
  `2026-07-07T15:30:22+02:00` — the offset makes the instant unambiguous), and the camera's own
  `timeZone` string **echoed back verbatim** (we correct the clock, never fight its tz/DST config).
  An unparseable camera reply counts as infinite drift → sync.
- **Visibility:** a sync after a big jump (>1h — the power-cut signature) logs at **warn**
  (persisted to app_logs); small corrections log info. Failures log warn.
- **Scope:** Hikvision only (`isClockSyncable` capability guard); the Dahua driver's CGI has no
  ISAPI time endpoint — a Dahua clock sync would be its own driver work.
- Rejected alternative: camera-side **NTP against the booth** (chrony on the appliance). More
  standard, but adds a provisioning dependency per booth and the camera polls NTP on ITS schedule
  — a freshly power-cycled camera could still sit at 1970 for a while, which is precisely the
  moment that matters.

## Camera PUSH — "Alarm Server" event notifications (2026-06-22)

Separate from the **pull** snapshot path above: newer Hikvision firmware can **push** an event to
us. Under **Event → Smart/VCA** (e.g. line crossing / intrusion / "Vehicle Detection") the unit
exposes **Detection Target: Human / Vehicle** — selecting **Vehicle** + **Notify Surveillance
Center**, then **Alarm Settings → Alarm Server**, makes the camera **HTTP-POST an
`EventNotificationAlert`** to a URL we host on each detection. Same machine-call shape as the
[[dingtian-relay]] Input Link push — no polling.

- **Ingress:** `POST /api/devices/hikvision/:deviceId/event` (`apps/server/src/routes/hikvision-alarm.ts`).
  **Source-IP guarded** (must come from the device's configured `host`) + **optional HTTP Digest**
  (some firmware can't authenticate the Alarm Server call → source-IP only). NOT behind the SPA
  cookie/CSRF (it's a device call), exactly like the Dingtian push.
- **Config:** added to the `hikvision` driver — `alarmPushEnabled` (bool), `pushUser`/`pushPassword`
  (optional Digest). The driver is now `pushesToBackend: true`, so first-run setup offers the backend
  push IP. Point the camera's Alarm Server at `http://<backend-ip>:<port>/api/devices/hikvision/<deviceId>/event`.
- **Discovery-first:** the endpoint is **permissive** — accepts ANY content-type as raw bytes (event
  XML, multipart-with-JPEG, or JSON; Hik's format varies by model/firmware), records the **verbatim
  body** as a `kind:"alarm"` device_event, and best-effort extracts `eventType` / `target` / `plate`
  / `dateTime` / `channelID`. The point of this first cut is to **see exactly what a given camera
  sends** (inspect via `GET /api/events` or the server log) before wiring it to the read bus.
- **Not yet a barrier trigger.** It records + breadcrumbs only; it does NOT emit a `DeviceReadEvent`
  or open anything. A plate read is **advisory, never the sole reason** a barrier opens
  ([[append-only-event-chain]], [[opencv-anpr-service]]). Two consumers were since designed off this
  same vehicle event — see **[[lane-presence-and-anpr-entry]]**: (a) BUILT — advisory lane busy/free
  booth lights; (b) BUILT (2026-06-22) — the ANPR "bridge" (`anpr-entry.ts`) that snapshots → ANPR →
  emits a `kind:"plate"` read for a SUBSCRIBER match through the existing gated flow (a small
  `apps/server` handler class, not a service). If the camera ever emits its own `<plateNumber>` we'd
  use it directly; this `DS-2CD1043G2`
  does not, so the server pulls the frame and hands it to the [[opencv-anpr-service|vision service]].

### "Subscribers auto-enter but don't auto-exit" — a 4-layer CAMERA fault, NOT our code (2026-06-27)

A long debugging session on the **DS-2CD1047G3H-LIU** exit camera (`10.0.10.13`, exit-lane). The
symptom: subscribers (e.g. Caca) auto-entered via ANPR fine but **never auto-exited**. **Every
assumption about *our* code was wrong; all four real causes were camera-side.** Method that finally
cracked it: a **dumb HTTP sink** (`scratch-camera-sink.py`) the camera's Alarm Server was pointed
at, to see — verbatim — what the camera actually sends, independent of our app's parsing/acceptance.

The wrong turns, and what was actually true:

1. **Wrong assumption: "the exit camera 503s, so harden the snapshot retry / reduce load."** The 503
   storm in the data was mostly **manual main-stream testing**: on this G3H, `channels/101/picture`
   (MAIN) **503s instantly every time** — structurally unavailable, not "busy" — while `102` (SUB)
   serves 10/10 rapid pulls cleanly. AND the camera **503s on *any* stream when its connection slots
   are exhausted** (a held web UI / live-view, parallel experiments); a **reboot clears stuck slots**.
   So the snapshot retry/flow code was fine. See [[#HTTP 503 "Device Busy"]] + the
   `g3h-main-stream-snapshot-503` memory. (The exit/subscription FLOW logic, camera→relay binding,
   and `stream:"2"` config were all correct the whole time — verified: clean entry/exit pairs before
   the test storm, and after the reboot.)

2. **The actual blocker #1 — the exit camera never POSTed at all.** `alarmPushEnabled=true` in our
   config, but `10.0.10.13` had sent **ZERO** alarms ever (entry cam `.12`: 1126). The sink received
   nothing from `.13`; our `/event` endpoint logged no rejections either → the camera wasn't sending.
   Cause found in the camera's own **Diagnose Information** dump: **`Main Db is broken` /
   `db_restore failed` / `Going to reset cfg`** — the camera's internal config DB (`ipc_db`) was
   **CORRUPT**, plus repeated reboots. A broken config DB means the event→linkage→push pipeline can't
   reliably read its own config, so it silently never POSTs. **Fix: factory-reset the camera** (rebuilds
   `ipc_db`), then reconfigure. (If corruption returns after a clean reset → failing flash → RMA.)

3. **Wrong assumption: "missing gateway/DNS blocks the push."** A documented Hikvision note says a
   gateway is needed even same-subnet — but here it was **inverted**: the WORKING cam `.12` has NO
   gateway/DNS; the broken `.13` HAD both. Red herring. Gateway/DNS was not the cause.

4. **The actual blocker #2 — after reset, the camera pushed PLAIN MOTION, not vehicle.** Post-reset
   `.13` POSTed `<eventType>VMD</eventType>` with **no target tag**. The backend gates on
   `target == "vehicle"` (`hikvision-alarm.ts` `isVehicleActive`, matches `<targetType>` /
   `<detectionTarget>` / `<objectType>`), so a plain-motion push is **ignored** → bridge never fires.
   The working `.12` sends `eventType=VMD` **with `target=vehicle`**. The difference is the AcuSense
   **Detection-Target = Vehicle** filter ON the Motion event — defaulted OFF after factory reset.
   **Fix: enable Vehicle target classification** on `.13`'s Motion Detection. Confirmed live: a real
   drive-through then POSTed `<eventType>VMD</eventType> … <targetType>vehicle</targetType>` — exactly
   what the backend needs. (So "VMD/Motion" IS the right event for this camera class; it's the *target
   filter* that matters, not switching to a different event type.)

**Takeaways:** (a) a camera that's silently not-pushing looks identical to "fine" in our logs — the
sink-to-prove-it-sends method is the fastest disambiguator; flagged as an observability gap (a
`alarmPushEnabled=true` camera with 0 pushes ever should be a surfaced condition, like the
[[device-status-monitoring|reader-liveness]] fix). (b) For ANPR the camera must send a **vehicle
`targetType`** — verify the push body, not just the UI toggles. (c) Hikvision config-DB corruption
is real; factory reset is the cure. None of this was a code bug. See
[[lane-presence-and-anpr-entry]] for the push→bridge→exit path.

### Gotchas learned the hard way (2026-06-22 field session)

Several traps surfaced trying to get a real camera to push. In order of how long each cost:

- **WSL rewrites the inbound source IP.** On the dev host (WSL mirrored mode), an inbound LAN packet
  arrives at our server with its **source rewritten to the host's own IP** (`10.0.10.203`), not the
  camera's. The source-IP guard then rejects every push as a mismatch. Fix: a per-device
  **`skipSourceIpCheck`** config flag (a Setup checkbox) that bypasses the IP guard — the signed
  ledger + optional Digest remain the real guards. Leave OFF on a normal LAN.
- **The setup checkbox saved booleans as the STRING `"true"`.** The generic config-field form had no
  boolean renderer, so a `type:"boolean"` field fell through to a text input. Fixed (checkbox
  renderer); the server also coerces `"true"`/`1`/`yes`/`on` defensively.
- **The camera's "Test" button proves almost nothing.** It does a TCP/connectivity probe and reports
  "service available" on ANY HTTP reply (even our 404) — it does **not** POST a real event to your
  URL. Only a real detection (or the ISAPI `httpHosts/<id>/test`) actually exercises the path.
- **`httpBroken` latches.** Once the camera marks the host broken (from earlier failed deliveries),
  it stays `true` across reboots and won't retry. Clear it by **re-PUTting** the httpHost config
  (`PUT /ISAPI/Event/notification/httpHosts/1` with `<httpBroken>false</httpBroken>`).
- **"Notify Surveillance Center" ≠ the HTTP Alarm Server** on some firmware (separate upload
  channels). Always confirm the **Arming Schedule** covers the test time, too (a silent killer).
- **⭐ THE ROOT CAUSE (2026-06-22): no detection AREA drawn.** This is what actually defeated us for
  most of a day. On the motion/smart-detection page there's a **Draw Area** step — if **no region is
  drawn on the frame, the camera detects nothing, generates NO event, and therefore posts nothing**
  anywhere (httpHost, FTP, alarm stream all stay silent because there's no event upstream). Enabling
  the detection + ticking Notify Surveillance Center is **not enough** — you must draw the region.
  Once an area was drawn, the very first vehicle produced a clean POST. **Check this FIRST.**

### Confirmed real payload (DS-2CD1043G2-LIU, V5.8.10, 2026-06-22)

What this camera actually POSTs on a motion event with a target — captured end-to-end:

- **`Content-Type: multipart/form-data; boundary=boundary`**, one XML part named `MoveDetection.xml`
  (`Content-Type: application/xml`). A real frame/JPEG *may* be attached as a second part on other
  event types — our endpoint stores the readable head; splitting an image part to `snapshots` is a
  forward step (not needed for plain motion).
- The XML is an `EventNotificationAlert` with the fields we care about:
  - `<eventType>VMD</eventType>` (Video Motion Detection) + `<eventState>active</eventState>`
  - **`<targetType>vehicle</targetType>`** — the camera classifies **vehicle vs human ON-DEVICE**.
    (Field is `targetType`, NOT `detectionTarget`.) This means simple presence + class comes for
    free, no vision model needed for that part.
  - `<targetInfo><targetRect>` with normalized `X/Y/width/height` (0–1) — the **bounding box**.
  - `<channelID>`, `<macAddress>` (provenance), `<dateTime>` — **but the dateTime is GARBAGE**
    (`2032-…`) because this unit's **RTC is dead** (see below); we use our own server receive time,
    never the camera's. (No `<plateNumber>` — this is a motion event, not an ANPR camera.)

### If a camera still won't push — diagnostics (read its OWN state)

Only after confirming the **detection area is drawn** + arming schedule covers now + Notify
Surveillance Center is on. These read the camera directly (no cooperation from our server):

1. **`netstat` on the camera (via SSH) while you trigger** — watch for an OUTBOUND line
   `cam:port → server:3000`. It appearing = the camera fired and is delivering (then check our
   `/api/devices/hikvision/alarms`). None = no event was generated (almost always: **no area drawn**).
2. **`GET /ISAPI/Event/notification/alertStream`** (Digest, needs a clean handshake) — the live event
   bus. NB: a `curl --digest` tap that fails the handshake returns empty and looks like "no events"
   — don't over-read silence here (this misled us); the netstat watch above is more reliable.
3. **SSH `showStatus` / `dmesg`** expose internal state. ⚠ **Caveat learned the hard way:** these
   surface scary-looking strings that are **red herrings** — `EventScribe: except`, a `diskfull`
   error on `Event/triggers` (on a camera with **no disk**), and `fh rtc get time error` / a 1970
   clock. On our unit ALL of these were present **and the camera worked fine** once an area was
   drawn. The dead RTC is real (hence the bogus `dateTime`) but **harmless** to event push. **Do NOT
   conclude "dead camera / RMA" from these** — they are not proof of a broken event engine.

> **Correction (2026-06-22):** an earlier version of this page concluded this DS-2CD1043G2-LIU was a
> **defective unit needing RMA**, based on the silent alertStream + `diskfull`/`EventScribe:except` +
> dead RTC surviving a full factory reset. **That was WRONG.** The camera was healthy; the real cause
> was simply **no detection area drawn**, so no event was ever generated. The `diskfull`/RTC findings
> were unrelated quirks (RTC genuinely dead, but it doesn't block event push). Lesson: don't
> escalate to "hardware fault" while a basic config precondition (the drawn region) is unmet — and
> treat vendor status-API error strings as unreliable. The pull + [[opencv-anpr-service|vision]] path
> remains a valid fallback, but it was not needed here.