julian
e4827c9651
Research note from the recognizer-options query. fast-alpr v0.4.0 (MIT) — a swappable YOLOv9-detector + CCT-OCR pipeline on ONNX Runtime, CPU-only and offline — fits the decided vision-service architecture and is MIT end-to-end (code + published weights), so the ANPR path may not need the scoped AGPL exception. Flags the open caveats: verify model-weight provenance, and benchmark AL-plate accuracy (default global vs. the 40+ country EU model). fast-alpr is plate-only, so the vehicle-verification job stays ours to build. Decision kept open. Updates opencv-anpr-service (new "Recognizer evaluation" section + licensing nuance), vision-service (open/next), index, log. Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
4.0 KiB
4.0 KiB
type, tags, sources, updated, status
| type | tags | sources | updated | status | |||||
|---|---|---|---|---|---|---|---|---|---|
| decision |
|
2026-06-15 | open |
Decision: Host-side Vision Service (ANPR + vehicle verification)
Taken 2026-06-15, as part of the business-layer build (session-model).
Decisions
- Build a host-side vision service (opencv-anpr-service) that does ANPR (plate → identity) and vehicle-attribute verification (anti-spoofing witness) on snapshots from ordinary Hikvision/Dahua cameras.
- It replaces the dedicated edge-AI lpr-camera as the recognition path: ordinary IP cam →
snapshot (
Snapshot.bytes, already pulled by the camera driver) → vision service → plate + vehicle. Removes the special LPR camera from the bom as a requirement (still allowed as an option). - Deployment: a separate local Python/OpenCV microservice on the appliance, called over localhost HTTP by the Node backend. Fully offline (offline-first); its own process and failure domain; the host falls back to the ticket path if it's unavailable.
- Licensing exception: AGPL components (e.g. YOLO plate/vehicle models, OpenALPR) are permitted inside this service only, because it's a separate process not linked into the app — the app stays strictly MIT/Apache/BSD. Amends standing-decisions.
- Recognition is advisory, evidence is authoritative. A read never single-handedly authorizes a paid/access barrier open; it flags for reconciliation and attaches (with the source image) to the signed append-only-event-chain entry. Low confidence → fallback, never strand a car (fail-state-safety).
Why
- Replace vs. edge-AI camera: host-side recognition on cheap IP cams shifts cost from per-lane smart cameras to one compute box + our software; gives us the raw image for the second job below.
- Vehicle verification is the real prize (user-driven, 2026-06-15): plate-only ANPR can't catch a printed/spoofed plate on a different car. Extracting vehicle attributes/fingerprint lets the system reconcile the car, not just the number — directly filling the independent-witness gap the append-only-event-chain calls out as unbuilt.
- Separate-process + AGPL-scoped keeps the app's permissive-license guarantee intact while not crippling accuracy (the strict permissive-only ANPR path is markedly weaker — that tradeoff was weighed and the scoped exception chosen).
Rejected / alternatives
- Strict permissive-only ANPR in-app — license-clean but weaker accuracy and more build; the separate-process AGPL exception was chosen instead.
- Keep the edge-AI LPR camera as primary — viable fallback if host-side accuracy disappoints; not chosen now, kept on the table in opencv-anpr-service.
- Embed OpenCV in Node (opencv4nodejs/WASM) — rejected: native-build pain, weaker model ecosystem, no process isolation, shares the app's failure + license surface.
Open / next
- Recognizer selection: fast-alpr (MIT, YOLOv9 detector + CCT OCR on ONNX Runtime) is the evaluated baseline as of 2026-06-19 — fits this decision's shape (separate offline localhost process, swappable) and is permissive end-to-end, so the ANPR path may not need the AGPL exception (pending a model-weight-provenance check). Plate-only: it does NOT cover the vehicle-verification job. Open: the provenance check + an accuracy benchmark on real Albanian plates (default global vs. the 40+ country EU model). See opencv-anpr-service "Recognizer evaluation".
- Vehicle-model selection + fingerprint method + anomaly threshold (the Job-2 anti-spoofing stage, still ours to build on the same ONNX runtime) (opencv-anpr-service).
- Appliance compute footprint (CPU vs. small GPU/NPU) — bom / open-questions.
- Service API + the Node-side adapter; per-camera opt-in wiring.
- Reconciliation logic that consumes plate+vehicle witness vs. commanded opens (still unbuilt — see append-only-event-chain, reconciliation).