Skip to content

Remote Roadside ANPR System

Remote Roadside ANPR, Vehicle & Person Detection Prototype Plan

Section titled “Remote Roadside ANPR, Vehicle & Person Detection Prototype Plan”

Project: Solar-Powered Remote Roadside ANPR, Vehicle & Person Detection Network Prototype
Location context: Residential neighbourhood in New Zealand
Document purpose: Practical prototype and implementation plan
Version: Final 3.0 — October 2026


  1. Introduction
  2. Problem Statement
  3. Prototype Goal
  4. Recommended Fastest Prototype Strategy
  5. System Architecture
  6. Prototype Phases
  7. Camera Recommendation
  8. Night-Time Number Plate Capture
  9. Existing Software Options
  10. Recommended Open-Source Route — Frigate
  11. Alternative Open-Source / DIY Software
  12. Commercial / Ready-Made ANPR Software
  13. Custom Python ANPR Pipeline
  14. Person Detection, Face Capture and Optional Face Recognition
  15. Remote Solar Camera Node
  16. Power System Design
  17. Battery Sizing
  18. Solar Panel Sizing
  19. 4G / LTE Connectivity
  20. Weatherproof Enclosure
  21. Local vs Remote Processing
  22. Networking and Security
  23. Database Design
  24. Common, Rare and New Vehicle Detection
  25. Incident Search Workflow
  26. Storage and Retention
  27. Privacy and Deployment Considerations
  28. Prototype Test Plan
  29. Success Criteria
  30. Suggested Hardware List
  31. Suggested Software Stack
  32. Recommended Repository Structure
  33. Step-by-Step First Prototype
  34. Second Prototype — Remote Solar Node
  35. Future Multi-Camera Architecture
  36. Features to Avoid Initially
  37. Risks and Mitigations
  38. Final Recommended Build Order
  39. Definition of Done
  40. Reference Links

Vehicle theft and suspicious vehicle activity have become a concern in the local residential area. A stolen vehicle may leave the area within minutes, while conventional household CCTV often records only a dark or blurred vehicle image at night.

The key problem is that ordinary wide-angle security cameras are normally designed to show the overall scene rather than capture a moving licence plate. At night, headlights, reflective number plates, motion blur and automatic camera exposure can make the plate unreadable even when the vehicle itself is visible.

This project proposes a small, privately operated ANPR prototype that can:

  • detect vehicles passing a defined road location;
  • capture readable number plates during the day and at night;
  • record the plate, time, camera location and direction of travel;
  • retain a supporting plate image and vehicle snapshot;
  • provide fast searching around the time of a theft or other incident;
  • distinguish frequently observed vehicles from rare or newly observed vehicles;
  • eventually support several remote camera nodes;
  • operate at locations where there is no mains power or fixed Internet connection.

The intended physical deployment is a small standalone camera node positioned beside a road, at a carefully selected fixed location where it can observe passing vehicles.

The preferred prototype location is:

  • on private property immediately beside or near the road;
  • close enough to obtain a useful plate angle and plate size;
  • positioned so that the camera looks at a narrow capture zone rather than the whole street;
  • physically secured against weather and tampering.

The prototype is not limited to being mounted on a house. A later node may be completely independent of a nearby building and may operate using:

ANPR camera
+ optional overview/person camera
+ solar panel
+ LiFePO4 battery
+ MPPT charger
+ 4G/LTE router
+ weatherproof locked enclosure

If a proposed position is on a council road reserve, berm, street-light pole, utility pole, footpath area, or other land that is not privately owned, permission from the relevant land/infrastructure owner must be confirmed before installation.

The first objective is not to build a neighbourhood-wide surveillance system.

The first objective is to prove one thing:

Can one camera reliably capture and recognise moving New Zealand licence plates at night?

Only after that has been demonstrated should the project expand to remote solar nodes, multiple cameras, analytics or a custom dashboard.


A normal CCTV camera may produce footage like:

Vehicle visible: YES
Vehicle colour: MAYBE
Driver visible: NO
Number plate readable: NO

The reasons commonly include:

  • motion blur;
  • long night-time exposure;
  • headlight glare;
  • plate overexposure;
  • insufficient infrared illumination;
  • plate occupying too few pixels;
  • camera positioned at too large an angle;
  • wide-angle lens;
  • excessive digital zoom.

Therefore, this project must solve the optical capture problem first and the OCR problem second.

Recommended priority:

Camera / illumination ~70%
Positioning / lens ~20%
Software / OCR ~10%

If the original frame does not contain readable characters, software cannot reliably reconstruct them.


The complete prototype should eventually perform:

Vehicle passes road
|
v
Camera captures plate
|
v
ANPR / OCR reads plate
|
v
Timestamp + direction + confidence
|
v
Store plate crop + vehicle image
|
v
Database
|
v
Search by incident time

Example event:

Time: 2026-10-04 01:17:21
Camera: NODE-01
Plate: ABC123
Confidence: 97%
Direction: Eastbound
Status: NEW
Plate image: ABC123_011721_plate.jpg
Vehicle: ABC123_011721_vehicle.jpg

The fastest route is not to build the full system from scratch.

Recommended order:

ANPR camera
|
record test footage
|
manually inspect plates

Goal:

Moving night-time plates must first be readable by a human.

Stage B — Use existing open-source ANPR software

Section titled “Stage B — Use existing open-source ANPR software”

Recommended first software:

Frigate NVR with built-in License Plate Recognition

This lets the project test:

RTSP camera
|
Frigate
|
vehicle detection
|
licence plate detection
|
OCR
|
history / search

without first writing the complete Python pipeline.

Once the camera has been proven and real NZ plate samples have been collected:

OpenCV
+ YOLO
+ PaddleOCR
+ custom NZ validation
+ SQLite/PostgreSQL

can replace or complement Frigate.

Only after Stage A/B works reliably.

This prevents spending money on:

  • batteries;
  • solar panels;
  • 4G routers;
  • cabinets;
  • poles;

before knowing whether the camera can actually read plates.


ROAD
================================================
Capture Zone
|
v
[ Vehicle ]
|
v
[ ANPR Camera ]
|
RTSP / LAN
|
v
[ Test PC ]
|
+-------+-------+
| |
v v
[ Frigate ] [ Raw Video ]
|
v
[ LPR ]
|
v
[ Events / Search ]

SOLAR PANEL
|
v
MPPT CONTROLLER
|
v
LiFePO4 BATTERY
|
+-----------+-----------+
| |
v v
ANPR CAMERA 4G ROUTER
| |
[optional overview camera] |
| |
+------- Ethernet ------+
|
v
4G / LTE
|
v
Internet
|
v
HOME / CLOUD SERVER
|
+-------------+-------------+
| |
v v
DATABASE DASHBOARD

Use one ANPR camera near a location where power and Ethernet are temporarily available.

Do not use solar yet.

Test:

  • daylight;
  • night;
  • headlights;
  • 30 km/h;
  • 50 km/h;
  • rain;
  • approaching vehicles;
  • departing vehicles.

Success:

Most moving plates are manually readable.


Install Frigate.

Connect the RTSP camera.

Enable LPR.

Success:

Camera -> Frigate -> recognised plate -> searchable history

Use the same saved footage.

Build:

Video
|
YOLO vehicle detection
|
plate detector
|
PaddleOCR
|
NZ validation
|
CSV / SQLite

Compare custom recognition against Frigate.


Move the camera to a location without mains power.

Add:

  • battery;
  • solar panel;
  • MPPT controller;
  • 4G router;
  • weatherproof enclosure.

Add Node 02, Node 03, etc.

Only do this after Node 01 has demonstrated acceptable reliability.


A strong candidate is:

Hikvision iDS-2CD7A26G0/P-IZHS — 8–32 mm

Why:

  • purpose-built ANPR functionality;
  • 2 MP progressive scan sensor;
  • DarkFighter low-light capability;
  • motorised 8–32 mm lens;
  • optical zoom;
  • built-in 850 nm IR;
  • manual / semi-auto focus;
  • shutter range to very fast exposures;
  • RTSP/network streaming;
  • weather-resistant body;
  • onboard ANPR capability;
  • suitable for narrow road capture zones.

For ANPR, a good 2 MP dedicated LPR camera can outperform a cheap 8 MP CCTV camera.

Example:

8 MP ordinary CCTV
-------------------
Street: excellent
Vehicle: excellent
Plate: white / blurred
2 MP ANPR camera
----------------
Street: dark
Vehicle: dark
Plate: ABC123

The second image is more useful for this project.


Night capture should use a dedicated plate capture zone.

Example:

CAPTURE ZONE
|
v
-------------------------------------------------
[ ABC123 ] ----->
-------------------------------------------------
/
/
/
[ ANPR Camera ]

Recommended initial settings:

Resolution: 1920 x 1080
Frame rate: 25–30 fps
Shutter: 1/500 second
IR: ON
IR wavelength: 850 nm
Gain: limited
WDR: test OFF initially
Focus: fixed after setup
Codec: H.264/H.265

If motion blur remains:

1/500
|
1/750
|
1/1000

Increase illumination rather than simply allowing longer exposure.

Aim for:

Horizontal angle: preferably < 30 degrees
Vertical angle: preferably < 30 degrees

Use optical zoom to make the plate occupy meaningful pixels.

Do not use one wide camera to perform both general surveillance and long-distance plate OCR.


There are three sensible software strategies.

OptionCostSetup SpeedCustomisationRecommended Use
FrigateFree/open sourceFastHighBest first prototype
OpenALPR open-sourceFreeMedium/HardHighExperiments / reference
Custom YOLO + PaddleOCRFreeSlowestMaximumPhase 2/3 development
Camera onboard ANPRIncluded with ANPR cameraVery fastMediumGreat hardware test
Commercial ANPR productsPaidFastVariesLater comparison

The fastest prototype should combine:

Dedicated ANPR camera
+
Frigate

For the initial prototype, enable vehicle and person detection where useful. Treat face detection/cropping as optional, and keep automatic facial identity matching outside the baseline scope.

while also testing the camera’s own ANPR feature if available.


10. Recommended Open-Source Route — Frigate

Section titled “10. Recommended Open-Source Route — Frigate”

Frigate is currently the best starting point for this prototype because it already provides:

  • IP camera / RTSP support;
  • video recording;
  • object detection;
  • event history;
  • snapshots;
  • retention rules;
  • MQTT integration;
  • built-in licence plate recognition;
  • known plate matching;
  • dedicated LPR camera mode.

Current Frigate LPR uses:

YOLOv9 licence plate detector
+
PaddleOCR

and can operate locally after the models have been downloaded.

It also supports dedicated LPR cameras with a narrow field of view.

Instead of initially building:

RTSP reader
vehicle tracking
plate detection
OCR
history
image storage
search UI

your first prototype can let Frigate provide most of those functions.

This means the first software experiment becomes:

1. Install Docker
2. Install Frigate
3. Add camera RTSP URL
4. Enable LPR
5. Define dedicated LPR camera
6. Pass vehicles
7. Check recognised plates

That could save a large amount of initial programming effort.


11. Alternative Open-Source / DIY Software

Section titled “11. Alternative Open-Source / DIY Software”

OpenALPR has an open-source library that can:

  • analyse images;
  • analyse video;
  • return plate text;
  • be called from C++, Python, C#, Java, Node.js and Go.

However, the open-source project is older and uses an AGPLv3 licence.

It is useful as:

  • a reference implementation;
  • a quick experiment;
  • a comparison against a custom pipeline.

It would not be my first choice for this prototype in 2026.


This is the recommended long-term custom route.

Pipeline:

RTSP / MP4
|
OpenCV
|
YOLO vehicle/person detector
|
YOLO plate detector
|
plate crop
|
PaddleOCR
|
NZ plate validation
|
deduplication
|
database

Benefits:

  • complete control;
  • can tune specifically for New Zealand plates;
  • can use your own training data;
  • can calculate familiarity/anomaly scores;
  • integrates easily with your own C# backend.

Disadvantage:

  • takes longer to build;
  • more tuning;
  • requires your own tracking and event logic.

Commercial solutions are worth testing after the open-source proof of concept.

Possible categories:

Some dedicated Hikvision and other professional ANPR cameras can perform recognition in the camera itself.

Advantages:

No GPU server required at remote site
Low bandwidth
Plate metadata produced at camera

This is especially useful for solar installations.

Commercial engines may provide:

  • better recognition accuracy;
  • plate-country models;
  • REST APIs;
  • RTSP stream processing;
  • support;
  • easier deployment.

For this project, the correct strategy is:

Do not buy expensive software before testing Frigate and the camera’s built-in ANPR.

A commercial engine can later be run against the same recorded test set and objectively compared.


Once the camera is proven, create a separate Python prototype.

Suggested components:

Python 3.11+
OpenCV
Ultralytics YOLO (vehicle/person/plate detection)
PaddleOCR
SQLite
FastAPI (optional)

Processing:

RTSP stream
|
frame sampling
|
vehicle detection
|
tracking
|
plate detection
|
best-frame selection
|
plate enhancement
|
OCR
|
normalisation
|
confidence scoring
|
deduplication
|
database

Common OCR confusion:

0 / O
1 / I
5 / S
8 / B

Store:

RawOCR
NormalizedPlate
Confidence

Never silently overwrite the raw OCR result.


14. Person Detection, Face Capture and Optional Face Recognition

Section titled “14. Person Detection, Face Capture and Optional Face Recognition”

The roadside system can also detect people. This should be designed in levels so that the first prototype remains useful without immediately creating a biometric identification database.

Recommended for the first prototype.

Person detected: YES
Timestamp: 01:14:28
Camera: NODE-01
Snapshot: person_011428.jpg

This allows an incident review to show that a person was present near a vehicle or capture zone.

Optional after person detection is working.

Person detected
|
Face visible?
|
save best face crop

The system may save the best available face image associated with an incident or person event.

The purpose at this level is evidence preservation, not automatic identity matching.

Level 3 — Face recognition / identity matching

Section titled “Level 3 — Face recognition / identity matching”

Not recommended as a Phase 1 requirement.

This level would compare a detected face with a stored biometric reference database and attempt to identify or match a person.

Example:

Face crop
|
face embedding / biometric template
|
reference database
|
possible match

This creates substantially higher privacy, security, governance and accuracy requirements.

In New Zealand, automated facial recognition used to identify or categorise people can fall within the Biometric Processing Privacy Code 2025. If this feature is ever considered, complete a privacy impact assessment and confirm necessity, proportionality, safeguards, transparency, retention and access controls before implementation.

For this project, the recommended rule is:

Detect and preserve useful imagery first; do not attempt automatic identity recognition in the initial prototype.


14.2 Why a separate overview camera may be better

Section titled “14.2 Why a separate overview camera may be better”

An ANPR camera is normally tuned for:

  • fast shutter;
  • narrow field of view;
  • infrared illumination;
  • reflective number plates;
  • plate readability.

That can produce an excellent plate image while producing a poor person image.

Recommended future roadside node:

ROAD
=====================================================
vehicle / person
|
+---------+---------+
| |
v v
ANPR / LPR camera Overview camera
narrow field wider field
fast shutter person/context
plate exposure low-light image
| |
+---------+---------+
|
v
Frigate
|
+-------------+-------------+
| | |
v v v
Plate Vehicle Person

For Prototype 1, one ANPR camera is enough.

For Prototype 2 or later, add an overview camera only if person/context imagery is genuinely useful.


Frigate can already detect person objects when an appropriate object-detection model is configured.

A custom Python pipeline can use:

YOLO
├── car
├── truck
├── motorcycle
└── person

Possible event record:

EventId
Timestamp
CameraId
VehiclePresent
Plate
PersonPresent
PersonSnapshotPath
FaceCropPath

Do not infer guilt from the presence of a person or vehicle. The output should be treated as investigative/search information only.


A remote node should be a standalone roadside unit intended for a fixed location beside the road, preferably on private land or another specifically authorised site.

It should not depend on a nearby house for mains power or fixed broadband.

A remote node should be a standalone unit.

Solar Panel
|
v
MPPT Controller
|
v
LiFePO4 Battery
|
DC Fuse / Breaker
|
+-----------+-----------+
| |
v v
ANPR Camera LTE Router
| |
+------ Ethernet -------+

Optional future addition:

Edge Computer
Raspberry Pi / Jetson / Mini PC

But do not add an edge computer initially unless required.

It consumes additional energy and adds maintenance complexity.


Keep the node DC-only.

Avoid:

12V battery
|
230V inverter
|
AC adapter
|
12V device

because every conversion wastes energy.

Prefer:

12.8V LiFePO4
|
+---- fused DC -> camera
|
+---- fused DC -> LTE router

A regulated DC/DC converter can be added where required.


Recommended chemistry:

LiFePO4

Reasons:

  • deep-cycle capable;
  • long cycle life;
  • stable voltage;
  • lighter than lead-acid;
  • high usable capacity;
  • suitable for solar applications.

Avoid using an ordinary car starting battery.

Conservative example:

ANPR camera 15 W
4G router 3 W
conversion loss 2 W
----------------------
Total 20 W

Daily energy:

20 W x 24 h
= 480 Wh/day

Nominal:

12.8 V x 100 Ah
= 1,280 Wh

Assuming ~80% usable:

~1,024 Wh

At 20 W:

1,024 / 20
= ~51 hours

So a 100 Ah battery provides roughly two days of autonomy in this simplified example.

Recommended:

Prototype minimum: 100 Ah
Better: 150 Ah
High resilience: 200 Ah

Actual sizing must be recalculated after measuring real camera and router consumption.

Use a battery with:

  • low-temperature charge protection;
  • suitable BMS;
  • ideally a protected enclosure.

Do not allow a LiFePO4 battery to be charged outside its manufacturer’s safe temperature range.


If consumption is approximately:

480 Wh/day

a nominal 100 W panel is too marginal for a reliable 24/7 node.

Example theoretical output:

100 W x 4 peak-sun-hours
= 400 Wh

and that is before real-world losses.

Recommended prototype sizing:

Minimum: ~200 W
Preferred: ~250–300 W
More robust: ~300–400 W

The final design should account for:

  • winter solar production;
  • cloudy days;
  • shading;
  • panel orientation;
  • charging efficiency;
  • battery efficiency;
  • dirt;
  • system losses.

Use an MPPT controller rather than a basic PWM controller.


Recommended class:

Industrial LTE router

Example:

Teltonika RUT241

Useful characteristics:

  • 9–30 V DC input;
  • Ethernet;
  • LTE;
  • external antenna support;
  • low power consumption;
  • suitable for unattended installations.

Published power figures are approximately around:

~1.3 W LTE idle
~1.5 W with one LAN device
~3.3 W under high traffic/load

Actual field consumption must be measured.

Do not continuously stream video over 4G unless necessary

Section titled “Do not continuously stream video over 4G unless necessary”

Preferred:

Camera records locally
|
ANPR event
|
plate metadata
|
small JPEG
|
4G upload

Example:

{
"camera": "NODE-01",
"time": "2026-10-04T01:17:21",
"plate": "ABC123",
"confidence": 0.97,
"direction": "eastbound"
}

Full video can remain on local SD storage and only be downloaded when an incident occurs.


Use a locked IP65/IP66 outdoor enclosure.

Inside:

+--------------------------------+
| Weatherproof Enclosure |
| |
| LiFePO4 battery |
| MPPT controller |
| fused distribution |
| DC/DC converter |
| LTE router |
| low-voltage protection |
| optional monitoring |
+--------------------------------+

Camera remains outside.

Solar panel remains outside.

LTE antenna may be mounted outside if required.

Add:

  • cable glands;
  • fused circuits;
  • surge protection where appropriate;
  • ventilation/thermal consideration;
  • tamper-resistant fasteners;
  • lock;
  • labelled isolation point.

Electrical work should follow the requirements applicable to the actual installation.


Best for initial solar node.

Camera
|
onboard ANPR
|
plate + JPEG
|
4G

Advantages:

  • lowest remote compute requirement;
  • lower power;
  • lower bandwidth;
  • simpler remote hardware.

Remote camera
|
4G RTSP
|
home server
|
Frigate / Python

Advantages:

  • centralised AI;
  • easy software updates.

Disadvantages:

  • higher data usage;
  • needs reliable 4G;
  • continuous RTSP consumes bandwidth.

Camera
|
Jetson / Mini PC
|
ANPR
|
metadata only

Advantages:

  • flexible;
  • low network traffic.

Disadvantages:

  • additional power;
  • complexity;
  • heat;
  • maintenance.

Prototype order:

1. camera onboard ANPR
2. Frigate at home/local test network
3. custom Python
4. edge computer only if justified

Remote cameras must not simply expose RTSP or web administration directly to the Internet.

Preferred:

Remote node
|
LTE router
|
VPN tunnel
|
Home network

Use:

  • WireGuard or equivalent VPN;
  • unique passwords;
  • HTTPS where available;
  • firewall rules;
  • no unnecessary inbound ports;
  • firmware updates;
  • separate camera VLAN at home;
  • restricted admin access.

Example home network:

Camera VLAN
192.168.30.0/24
ANPR server
192.168.30.10

Camera
------
Id
Name
Description
Direction
Enabled
CreatedAt
VehicleSighting
---------------
Id
CameraId
TimestampUtc
PlateRaw
PlateNormalized
PlateConfidence
VehicleType
VehicleColour
Direction
PlateImagePath
VehicleImagePath
CreatedAt
PlateStatistics
---------------
PlateNormalized
FirstSeen
LastSeen
SeenCount
NightSeenCount
DaySeenCount
TypicalHour
TypicalDirection
UpdatedAt
Incident
--------
Id
Title
StartTime
EndTime
Description
CreatedAt

24. Common, Rare and New Vehicle Detection

Section titled “24. Common, Rare and New Vehicle Detection”

Do not discard common vehicles.

Record sightings and classify familiarity.

Example:

ABC123
Seen: 187
Typical: 07:20–08:10
Direction: Eastbound
Status: COMMON

Another:

XYZ999
Seen: 1
First seen: 01:17
Status: NEW

A later anomaly score might use:

First ever appearance +40
Seen fewer than 3 times +25
Seen midnight–5am +20
Unusual direction +10
Outside normal time pattern +10
Near incident time +20

This is only a search prioritisation tool.

It must not be interpreted as evidence that a driver has committed an offence.


If a vehicle theft occurs around 01:10:

Incident:
00:45–01:30

System query:

Show all plates
between 00:45 and 01:30
from Camera 1 / Camera 2
sorted by rarity

Example:

01:17 XYZ999 NEW 97%
01:13 KLM456 RARE 95%
01:08 ABC123 COMMON 98%

Then inspect:

  • vehicle image;
  • plate image;
  • direction;
  • other cameras;
  • full local footage.

Suggested prototype policy:

Full video: 48–72 hours
Vehicle snapshots: 14 days
Plate crops: 30 days
Plate metadata: 30 days
Incident evidence: preserved manually

These are engineering starting points, not legal requirements.

Implement automatic deletion.

If an incident occurs:

Create Incident
|
Select time window
|
Preserve linked media

Use conservative privacy practices.

Principles:

  • capture only what is needed for security;
  • avoid viewing into neighbours’ private areas;
  • avoid recording audio unless genuinely necessary;
  • do not publish plate histories;
  • restrict system access;
  • use secure passwords and VPN access;
  • limit retention;
  • preserve relevant footage for Police rather than public identification.

For the prototype, the safest practical approach is to mount the camera on private property.

Do not install poles, cabinets, solar equipment or cameras on council road reserve, berms, street furniture or other land without confirming permission from the relevant owner/authority.


Build a repeatable dataset.

Target:

100–200 vehicle passes

Include:

ConditionTest
Dayslow vehicle
Day30 km/h
Day50 km/h
Nightstreet lighting
Nightheadlights toward camera
Nightrear plate
Nightwet road
Weatherrain
Traffictwo vehicles close together
Specialpersonalised plate
Specialdirty plate
Specialmotorcycle if visible

For every sample record:

Was plate manually readable?
Was plate detected?
Was OCR correct?
Confidence?
Vehicle direction correct?

Initial engineering targets:

Readable plate capture: >= 95%
Correct OCR: >= 90%
Readable plate capture: >= 90%
Correct OCR: >= 85%

The important distinction:

Capture failure != OCR failure

Track both separately.


ItemRecommendation
CameraHikvision iDS-2CD7A26G0/P-IZHS 8–32 mm or equivalent dedicated ANPR camera
NetworkEthernet
Powertemporary mains / PoE for testing
Computerexisting Windows/Linux PC
Storageexisting PC storage
SoftwareFrigate first
DevelopmentPython later

Do not buy solar hardware yet.


ItemRecommended class
CameraDedicated 8–32 mm ANPR camera
SD card256–512 GB high-endurance
RouterTeltonika RUT241 or equivalent
Battery12.8 V 100–150 Ah LiFePO4
Solar~250–300 W starting design
ControllerMPPT 20–30 A class
Enclosurelockable IP65/IP66
ProtectionDC fuses / breakers
NetworkLTE SIM + VPN
Overview/person cameraOptional later; not required for first ANPR test
Edge computerNone initially

Final electrical sizing must be checked against actual measured loads and equipment specifications.


Docker
Frigate
SQLite / Frigate internal storage
Python
OpenCV
Ultralytics YOLO (vehicle/person/plate detection)
PaddleOCR
SQLite
FastAPI
Python ANPR worker
|
REST / message queue
|
ASP.NET Core API
|
PostgreSQL
|
Vue / Angular frontend

Do not start with the production architecture.


anpr-prototype/
│
├── README.md
├── docker-compose.yml
│
├── frigate/
│ └── config.yml
│
├── python/
│ ├── requirements.txt
│ └── app/
│ ├── capture/
│ │ └── rtsp_capture.py
│ ├── detection/
│ │ ├── vehicle_detector.py
│ │ └── plate_detector.py
│ ├── ocr/
│ │ └── plate_ocr.py
│ ├── processing/
│ │ ├── image_enhancement.py
│ │ ├── plate_normalizer.py
│ │ └── deduplicator.py
│ └── main.py
│
├── data/
│ ├── test-videos/
│ ├── snapshots/
│ ├── plates/
│ └── ground-truth.csv
│
└── docs/

This is the recommended quickest route.

Step 1 — Buy/borrow one proper ANPR camera

Section titled “Step 1 — Buy/borrow one proper ANPR camera”

Do not buy the entire remote system.

Use one dedicated ANPR-capable camera.


Mount on private property.

Temporarily provide:

PoE / mains
Ethernet

Aim at one narrow road zone.


Before installing any OCR:

Test at night.

Tune:

zoom
focus
shutter
gain
IR
camera angle
capture distance

Stop here until plates are readable.


Create your first benchmark dataset.

Example:

day_30kmh.mp4
night_30kmh.mp4
night_headlights.mp4
rain_night.mp4

Install using Docker.

Add the camera’s RTSP stream.

First verify:

Live view works
Recording works
Car detection works

Enable Frigate’s licence plate recognition.

For a dedicated plate camera, configure dedicated LPR mode.

Use saved debug plate images while tuning.


Create:

timestamp,actual_plate,frigate_plate,correct
21:03:11,ABC123,ABC123,true
21:05:42,XYZ987,XYZ987,true
21:07:01,DEF456,DEFG56,false

This creates objective accuracy numbers.


Once plate recognition is working, enable person detection in Frigate if the camera view includes an area where people may be relevant to an incident.

Verify:

person detected
snapshot saved
timestamp correct

Do not add face recognition at this stage.

If useful imagery of people cannot be obtained from the ANPR camera because the field of view is too narrow or exposure is optimised for reflective plates, plan an optional overview camera for the later remote node.


If the chosen camera supports onboard ANPR:

compare:

Frigate result
vs
Camera result

on the same vehicles.


Do not build the whole backend.

First Python goal:

Terminal window
python recognize.py night_test.mp4

Output:

timestamp,plate,confidence

Use:

YOLO
PaddleOCR

Choose based on measured results.

Keep Frigate and add your own API/analytics later.

Let the camera recognise plates and send events centrally.

Build/fine-tune the custom Python model.

This decision should be evidence-based.


34. Second Prototype — Remote Solar Node

Section titled “34. Second Prototype — Remote Solar Node”

Only after the first prototype passes.

Build:

solar
|
MPPT
|
battery
|
+------------+
| |
camera LTE router
| |
+--- LAN ----+

Then test:

  • voltage at sunrise;
  • voltage at sunset;
  • daily Wh consumption;
  • battery state of charge;
  • solar harvest;
  • cloudy-day performance.
  • LTE signal;
  • connection reliability;
  • VPN recovery;
  • remote reboot.
  • same image quality as local test;
  • local SD recording;
  • ANPR event upload.

Run for at least several real weather cycles before adding a second node.


Future:

NODE-01 --------\
NODE-02 ---------\
NODE-03 ----------> Central API -> PostgreSQL -> Dashboard
NODE-04 ---------/

Each event:

{
"nodeId": "NODE-02",
"timestamp": "...",
"plate": "ABC123",
"confidence": 0.97,
"direction": "westbound"
}

This allows:

  • same plate seen at different locations;
  • time correlation;
  • direction reconstruction;
  • rare/new vehicle analysis.

Do not implement route profiling until there is a clear lawful and privacy-conscious need.


Do not start with:

  • facial recognition;
  • automatic identity lookup;
  • public plate database;
  • neighbourhood-wide deployment;
  • cloud microservices;
  • Kubernetes;
  • gRPC;
  • custom mobile app;
  • vehicle make/model AI;
  • GPS tracking;
  • complex anomaly AI;
  • custom circuit boards.

They do not help answer the first engineering question.


Mitigation:

Better optics
Fast shutter
850 nm IR
Narrow capture zone
Smaller angle

Mitigation:

Measure real power
Oversize battery
Oversize solar
MPPT
Low-voltage cutoff
Remote battery telemetry

Mitigation:

Local recording
Metadata-only uploads
JPEG event upload
Download full video only after incidents

Mitigation:

  • mount out of easy reach;
  • locked metal enclosure;
  • tamper-resistant fasteners;
  • hidden identification;
  • offline alert;
  • optional tracker later.

Mitigation:

  • retain original plate image;
  • store confidence;
  • preserve raw OCR;
  • require manual verification before treating any plate as relevant.

Mitigation:

  • narrow capture zone;
  • short retention;
  • restricted access;
  • no public publication;
  • use for security incidents;
  • confirm permissions for any non-private-property installation.

Recommended sequence:

1. Buy ONE proper ANPR camera
|
2. Temporary powered installation
|
3. Tune nighttime plate capture
|
4. Record benchmark videos
|
5. Install Frigate
|
6. Test built-in Frigate LPR
|
7. Test camera-native ANPR
|
8. Enable/test person detection
|
9. Measure recognition accuracy
|
10. Build minimal Python comparison
|
11. Choose recognition engine
|
12. Build roadside solar + battery node
|
13. Add LTE + VPN
|
14. Run Node-01 for several weeks
|
15. Add optional overview/person camera if justified
|
16. Add database / anomaly analytics
|
17. Consider face cropping only if useful
|
18. Consider biometric face matching only after privacy/legal review
|
19. Only then consider more roadside nodes

This is deliberately different from trying to build everything at once.

It gives the fastest answer to:

Does this idea actually work under real local night-time road conditions?


The first complete prototype is successful when:

  • a moving vehicle enters the capture zone;
  • the plate is physically readable in the camera image;
  • the software detects it;
  • OCR produces the correct plate;
  • duplicate frames are collapsed into one vehicle event;
  • the event receives a timestamp;
  • direction is recorded if available;
  • a plate crop is saved;
  • a vehicle image is saved;
  • the sighting can be found by time;
  • night performance is acceptable;
  • person detection can save a timestamped person snapshot where the camera view supports it;
  • automatic face identity recognition is not required for Prototype 1;
  • data is automatically expired according to retention settings.

The remote prototype is successful when all the above works while:

  • powered only by battery/solar;
  • connected only by LTE/4G;
  • surviving multiple days of normal weather;
  • remotely accessible through a secure network connection;
  • recovering automatically after network or power interruptions.

These references should be rechecked before purchasing hardware because software and product specifications can change.

Frigate currently supports local LPR, dedicated LPR camera configuration, recognised plate fields, history filtering and MQTT events.

Useful specifications include the 8–32 mm lens option, 850 nm IR and dedicated licence-plate recognition support.

The open-source library is licensed under AGPLv3. Review licence requirements carefully before incorporating it into any commercial product.

The Biometric Processing Privacy Code is especially relevant if the system later performs automated facial identity matching or biometric categorisation.


For the very first experiment, build only this before placing anything at a remote roadside location:

Hikvision ANPR camera
|
Ethernet
|
Existing PC
|
Frigate

No solar.

No 4G.

No custom dashboard.

No C# backend.

No Raspberry Pi.

First prove:

Night moving vehicle
|
clear plate image
|
correct plate recognition

Once this works, the second experiment is a single authorised roadside node:

Roadside ANPR camera
|
optional overview/person camera
|
LiFePO4 + Solar
|
LTE router
|
VPN / remote access

Only after the remote node is stable should the project expand into a multi-location neighbourhood system.

That approach provides the fastest, lowest-risk path from idea to a working ANPR prototype.