1 · System overview
A camera/clip feeds the edge-detector; everything else talks over the Solace event mesh. toll-bridge is the ONLY writer of the stores; query-api only reads + publishes; the dashboard is the single public entry point.
YOLO26 + fast-alpr
+ colour · body-type
1 event / vehicle"] IOS["edge-detector-iOS
iPhone on-device
(same pipeline)"]:::opt CI["camera-ingestor
profile atlantic (optional)"]:::opt end BR{{"Solace PubSub+ — event mesh (reused)"}}:::reuse subgraph SVC["car-identify services (new)"] direction LR TB["toll-bridge
SOLE writer
dedup · bill · match"] VLM["vlm-worker
profile ai
exception enrich"] QA["query-api
read · approve · registry"] DASH["toll-dashboard
nginx: SPA + /api + /plates"] end subgraph STORE["Stores (new)"] direction LR CH[("ClickHouse
events + revenue MV")] MIN[("MinIO
plate snapshots")] end subgraph REUSE["Reused / external"] direction LR DT[("Eclipse Ditto
registry + twin")]:::reuse KC[("Keycloak")]:::reuse QVL[("Qwen3-VL
DashScope")]:::reuse end Cam --> ED Op --> DASH ED -->|"detected / exception"| BR ED -. "device plane" .-> CI IOS -. "device plane" .-> CI CI -. "republish toll" .-> BR ED -->|"snapshot"| QA BR --> TB BR -->|"exception"| VLM VLM -->|"enriched"| BR VLM -->|"read crop"| MIN VLM -->|"re-read"| QVL TB -->|"write"| CH TB -->|"write twin"| DT QA -->|"read"| CH QA -->|"snapshot"| MIN QA -->|"read registry"| DT QA -->|"approved / registry"| BR DASH -->|"/api"| QA DASH -->|"/plates"| MIN DASH -->|"OIDC"| KC QA -->|"verify JWT"| KC classDef opt stroke-dasharray:5 5,fill:#eef3f8,stroke:#90a4b8,color:#333; classDef reuse fill:#eef3f8,stroke:#7b8a9c,color:#222; classDef actor fill:#e7ecf2,stroke:#5e6b78,color:#222; style ED fill:#ffe9c7,stroke:#e0922f,color:#222 style TB fill:#ffe9c7,stroke:#e0922f,color:#222 style VLM fill:#ffe9c7,stroke:#e0922f,color:#222 style QA fill:#ffe9c7,stroke:#e0922f,color:#222 style DASH fill:#ffe9c7,stroke:#e0922f,color:#222 style CH fill:#ffe9c7,stroke:#e0922f,color:#222 style MIN fill:#ffe9c7,stroke:#e0922f,color:#222
Single-writer invariant: only toll-bridge writes ClickHouse + Ditto. vlm-worker (profile ai) handles only the exception branch; camera-ingestor (profile atlantic, dashed) is an optional pre-hop, off by default. The only direct browser→service HTTP call is dashboard→query-api (same origin via nginx).
2 · Detection → billing (happy path)
High-confidence plate (≥ 80%): detected once per vehicle, billed exactly $3, on the dashboard in under ~2s.
Dedup by tracking-ID + line crossing → one toll event per vehicle (no 30× over-billing). Images are referenced (image_ref), never streamed through the broker.
3 · Exception path (low-confidence → VLM → approve)
Below 80%, the plate is queued, re-read by the Qwen3-VL worker, and one-click approved by an operator. Fee is applied exactly once — at detected OR at approved, never both.
confidence"} DET -->|"≥ 80%"| HI["detected/high
status=billed · fee $3"] DET -->|"below 80%"| LO["exception/low
status=pending · fee 0"] LO --> VLM["vlm-worker
Qwen3-VL re-read
+ make/model/color"] VLM -->|"enriched"| EP["enriched_pending
fee 0"] EP --> OP([Operator reviews on dashboard]) OP -->|"approve via Keycloak"| AP["query-api publishes approved"] AP --> BILL["toll-bridge → approved
fee $3 set ONCE here"] HI --> CH[("ClickHouse")] BILL --> CH style VLM fill:#ffe9c7,stroke:#e0922f,color:#222
vlm-worker subscribes only the exception topic; it enriches + republishes enriched and never writes the stores. Operator identity comes from the validated JWT, never the request body.
4 · Vehicle registry + match ★
A registry of known vehicles (ETC / account holders, blacklist, unpaid) lives in Ditto. Every detection is matched against it — turning "count + bill" into real ANPR. Edits respect the single-writer rule: they are published, then applied by toll-bridge.
Registered Vehicles"] DASH -->|"POST /api/toll/vehicles"| QA["query-api"] QA -->|"publish registry/upsert"| BR{{"Solace"}} BR --> TB["toll-bridge
sole Ditto writer"] TB -->|"PUT attributes.registration"| DT[("Ditto
toll:vehicle:PLATE")] end subgraph MATCH["Match on EVERY detection"] direction LR DET["detection (plate)"] --> TB2["toll-bridge"] TB2 -->|"GET registration (30s cache)"| DT DT -->|"hit / miss"| TB2 TB2 --> ROW[("ClickHouse row
registered · owner · reg_status
+ make/model/color")] end style TB fill:#ffe9c7,stroke:#e0922f,color:#222 style TB2 fill:#ffe9c7,stroke:#e0922f,color:#222 style DT fill:#ffe9c7,stroke:#e0922f,color:#222
Best-effort, fail-open: no match → row still inserted + billed as an anonymous vehicle (registered=0); a Ditto outage never blocks revenue. Blacklist / unpaid are flagged on the console.
5 · Billing state machine (fee-once invariant)
The lynchpin of both billing correctness and the constant-time revenue view: a positive fee is written exactly once per track_id.
Stored in toll_events (ReplacingMergeTree by gate_id, track_id); the toll_revenue_agg materialized view answers SUM(fee) in sub-second over billions of rows.
6 · Three stores — one job each
Structured data, binary blobs and live state are different problems, so they live in different engines. The row only keeps a pointer to the image.
structured rows
SUM(fee) · filter · sort")] TB --> DT[("Ditto
current state
registry + twin")] QA["query-api / edge"] --> MIN[("MinIO
JPEG blobs · TTL 7d")] CH -. "image_ref → object key" .-> MIN style CH fill:#ffe9c7,stroke:#e0922f,color:#222 style MIN fill:#ffe9c7,stroke:#e0922f,color:#222 style DT fill:#ffe9c7,stroke:#e0922f,color:#222
ClickHouse = ledger (numbers to aggregate). MinIO = filing cabinet (images; the browser loads them directly via /plates). Ditto = digital twin (mutable state). Putting blobs in the DB would kill OLAP speed.
7 · Topic taxonomy (Solace)
All toll traffic lives under one domain; toll-bridge subscribes the whole subtree, each other service a narrow slice.
subscribes everything"] BR -->|"exception only"| VLM style TB fill:#ffe9c7,stroke:#e0922f,color:#222
Shape: STYL/iot/{domain}/{entityType}/{entityId}/v1/{function}/{action}. Registry action = upsert | delete. Isolation from the other demos is by the toll domain + unique Solace client names.
8 · Atlantic device-plane integration ★ optional
To let Atlantic manage the camera like any sensor, the edge can publish to the tenant-less ingress plane; a camera-ingestor maps camera→demo and republishes onto the toll domain — toll-bridge is unchanged. Default stays standalone (PUBLISH_PLANE=toll).
PUBLISH_PLANE=device"] -->|"STYL/iot/device/image-camera/{camId}/..."| CI["camera-ingestor
compose --profile atlantic"] MAP[("device → demo mapping
local cfg / atlantic api")] -.->|"camId → domain + gate"| CI CI -->|"STYL/iot/toll/gate/{gate}/..."| TB["toll-bridge (unchanged)"] CI -->|"unmapped"| DROP["drop + count
never guess a domain"] style CI fill:#ffe9c7,stroke:#e0922f,color:#222 style MAP fill:#ffe9c7,stroke:#e0922f,color:#222
This is the camera-ingestor anticipated by Atlantic's topic-taxonomy. Demo runs from a local mapping; "Atlantic truly owns it" additionally needs the image-camera device + mapping registered in atlantic-super-admin (future).
9 · Auth flow
Reuses the shared Keycloak realm (no custom auth). Viewing the dashboard requires login; the approve + registry-edit actions require a valid token.
Client vehicle-identification (public PKCE). No specific role required — any signed-in user can approve/register; the action is audited with their username.
10 · Deployment (docker-compose on the VM)
Single published port (atlantic pattern): the dashboard's nginx fronts everything. toll-bridge + query-api join the monorepo's network to reach Ditto. ClickHouse data + MinIO blobs persist on named volumes across the daily restart.
vehicle-identification-uat"] --> DASH["car-identify-dashboard :8095
nginx (only published port)"] DASH -->|"/api"| QA["car-identify-query-api
internal"] DASH -->|"/plates"| MIN[("car-identify-minio
internal")] QA --> CH[("car-identify-clickhouse")] TB["car-identify-toll-bridge"] --> CH QA --> MIN subgraph EXT["monorepo backend network (reused, external)"] DTN["ditto-nginx → Eclipse Ditto"] end TB -. "join network" .-> DTN QA -. "join network" .-> DTN VLM["vlm-worker · profile ai"]:::opt CI["camera-ingestor · profile atlantic"]:::opt classDef opt stroke-dasharray:5 5,fill:#eef3f8,stroke:#90a4b8,color:#333; style DASH fill:#ffe9c7,stroke:#e0922f,color:#222 style QA fill:#ffe9c7,stroke:#e0922f,color:#222 style TB fill:#ffe9c7,stroke:#e0922f,color:#222 style CH fill:#ffe9c7,stroke:#e0922f,color:#222 style MIN fill:#ffe9c7,stroke:#e0922f,color:#222
Containers are restart: unless-stopped + the Docker daemon is enabled on boot → the stack self-recovers after the daily VM stop/start. Schema migrations (e.g. registry columns) are applied once and persist in the volume.
11 · Camera & device options — Singapore cost comparison Jun 2026
The platform is camera-agnostic: it only needs one toll.frame.v1 event per vehicle on Solace. So the "camera" can be anything from a webcam you already own to a tolling-grade ANPR unit. Four routes, cheapest→richest. Plate text, colour and body-type come from our edge software (YOLO26 + fast-alpr + OpenCV) for routes A–C; make/model is added by the Qwen3-VL worker on hard cases. Smart ANPR cameras (route D) read the plate on-camera but still need our pipeline for make/model and a small camera-ingestor-style adapter to publish onto Solace.
| Approach | Example hardware (SG) | AI runs | Plate (ANPR) | Colour · type · make/model | Approx price (SGD, ex-GST) | Fit / notes |
|---|---|---|---|---|---|---|
| A · Reuse cameras you already have (software only) — cheapest | ||||||
| Existing CCTV/IP camera via RTSP + our edge on a spare PC/laptop | Any installed ONVIF/RTSP cam + reused PC | Off-camera (our service) | ✓ fast-alpr | ✓ colour+type (edge) · make/model via VLM | ~S$0* | Fastest PoC on installed gear. *uses hardware you own; one PC can serve several cameras. |
| B · Build new — ordinary IP camera + a compute box | ||||||
| 4MP IP bullet + reuse a PC | Hikvision DS-2CD2043G2 AcuSense / Dahua equiv. | Off-camera (our service) | ✓ fast-alpr | ✓ colour+type · make/model via VLM | ~S$100–180 | Camera only; long-range/IR plate optics (varifocal) ~S$200–400. |
| 4MP IP bullet + edge AI box (no server, 24/7) | IP bullet + NVIDIA Jetson Orin Nano Super 8GB | On the Jetson (edge) | ✓ fast-alpr | ✓ colour+type · make/model via VLM | ~S$500–580 | Jetson dev kit ≈ S$399. Self-contained lane unit; GPU runs YOLO26 comfortably. |
| 4MP IP bullet + low-cost edge box | IP bullet + Raspberry Pi 5 8GB + AI HAT+ (Hailo) | On the Pi (edge) | ✓ fast-alpr | ✓ colour+type · make/model via VLM | ~S$350–430 | Pi 5 ≈ S$120 + AI HAT+ ≈ S$130. Cheapest dedicated 24/7 edge. |
| C · iPhone as an all-in-one edge (our TollEdge app) ★ | ||||||
| iPhone = camera + compute in one (on-device) | Reuse any recent iPhone, or a new iPhone 16 | On-device (Neural Engine) | ✓ Apple Vision OCR | ✓ colour+type on-device · make/model via VLM | S$0 reused · S$1,299 new | App is free (free Apple ID for the demo). Zero rig, "AI in your pocket". See edge-detector-ios/. |
| D · Smart ANPR camera — plate read on-camera | ||||||
| Hikvision LPR/ANPR bullet | Hikvision DS-2CD7A26G0-P-IZHS8 | On-camera (plate) | ✓ on-camera ≥98% | ~colour/type (some models) · make/model via our VLM | ~S$1,288 | Needs a small adapter to publish plate→Solace. Best plate capture at speed/glare/night. |
| Hikvision ANPR (motorised varifocal) | Hikvision DS-TCG405-E (4MP LPR) | On-camera (plate) | ✓ on-camera | ~colour/type · make/model via our VLM | ~S$1,180 | Same integration as above. Varifocal for flexible lane distance. |
| Dahua ITC ANPR | Dahua ITC413 / ITC415 series | On-camera (plate) | ✓ on-camera 95–98% | ~colour/type · make/model via our VLM | ~S$1,000–1,900 (quote) | Distributor-quoted in SG; multilingual plate engine. |
| E · Enterprise / tolling-grade ANPR (free-flow, highway speed) | ||||||
| Dedicated tolling ANPR per lane | Survision · TagMaster · Tattile · Genetec-certified | On-camera (plate, certified) | ✓✓ highway-speed, free-flow | varies by vendor · make/model via our VLM | ~S$3,000–8,000+ /lane (POA) | Production highway gantries; overkill for the demo. Quote-only. |
Recommendation. Demo: route A (reuse a camera) or C (iPhone) — ~S$0. A clean self-contained lane unit: route B with a Jetson or Pi 5+AI HAT. Production highway: route E. Whatever the camera, the back end (Solace → toll-bridge → ClickHouse → dashboard) is identical.
Prices are indicative, ex-GST, as of Jun 2026 and vary by distributor/lens/license; enterprise ANPR is quote-only (POA). Sources:
Hikvision LPR SGD — sgcctv.com.sg ·
Jetson Orin Nano Super SGD — Ban Leong / Amazon.sg (US$249 list) ·
iPhone 16 SGD — Apple SG ·
Raspberry Pi 5 / AI HAT — Cytron SG.
Make/model/colour: see §2–§4 (edge colour+body-type on every vehicle; Qwen3-VL make/model on exceptions).