GoGhostGOGHOST
← Back to blog

How ALPR Works: 5 Steps From Plate Scan to Alerts and Privacy

Hand-drawn ALPR camera and plate details

Automated license plate recognition (ALPR, also called ANPR or LPR) uses cameras and software to capture a vehicle image, read the plate with optical character recognition, compare the result against databases such as hot lists, and trigger an alert or record when a match or rule applies. The whole sequence usually runs automatically, with no one touching a keyboard.

TL;DR:

  • ALPR reads carry timestamps, GPS positions, vehicle types, and colors, enabling stored records to reconstruct a vehicle’s movements across multiple locations.
  • Heavy fog or frost can sharply reduce plate detection, while mounting height, infrared lighting, lens choice, and regular cleaning improve conditions operators can control.
  • Before enforcement action, an operator should confirm the characters, issuing state, and current alert reason; a read alone is not proof.
  • State policies vary, with audit and retention rules; some states cap retention at three years, while copied hot lists can lag behind live records.
Goghost
Explore a Plate Privacy Option
GoGhost offers electric and mechanical license plate covers for drivers concerned about privacy on the road.
Explore GoGhost covers

Table of Contents

What happens between the camera flash and the alert

Once a vehicle enters the field of view, the system moves through a fixed chain of steps. NIJ-style operational guidance lists the core algorithmic stages, and the hardware side of the chain looks like this in practice:

  • Image capture: an overview image of the whole vehicle plus a tight plate close-up, usually taken with infrared illuminators and wide dynamic range (WDR) so glare and shadows do not wash out the characters.
  • Plate detection: software finds the rectangle that looks like a plate and narrows the region of interest, so later steps ignore bumpers, stickers, and signs.
  • Normalization: the cropped plate is rotated, resized, and corrected for skew and uneven lighting so it resembles a flat, head-on view.
  • Character segmentation: the plate is split into individual characters so each can be read on its own.
  • Matching and triggers: the text is compared against hot lists, either stored locally or queried from the cloud, and a match can fire an alert or log an entry.

Congressional summaries add that reads carry metadata such as timestamp, GPS position, vehicle type, and color, which is what makes later movement tracking possible.

Fixed, mobile, and portable setups

Where a camera sits changes what it can reliably read. Common setups differ in a few practical ways:

  • Fixed units on poles, gantries, or overpasses watch one lane or intersection continuously and are tuned to a known distance and angle.
  • Vehicle-mounted units ride on patrol cars or enforcement vehicles and read plates of parked and passing traffic while moving.
  • Portable or trailer units can be dropped at a work zone, event, or problem location and moved when the need passes.

Most systems pair a color overview camera with a dedicated infrared plate camera. The overview camera records context, such as make and color, while the plate camera is optimized purely for contrast on reflective plates, a split described in operational guidance. If you want to see how camera types differ in the field, our breakdown of the traffic camera types in the U.S. covers the detection side.

Pro Tip: Mounting angle matters as much as the camera itself. A steep angle or a lens with too wide a field of view shrinks characters in the frame and drags read rates down.

How software turns pixels into text

The software side repeats the same logic at a finer scale. Operational guidance names the stages: plate localization, orientation and sizing, normalization, segmentation, OCR, and syntactical and geometrical checks.

  • Localization finds the plate using object-detection models, which have largely replaced older edge-and-shape heuristics.
  • Orientation correction straightens the crop so the OCR step sees consistent input.
  • Segmentation and OCR isolate characters and classify them, either by template matching against stored glyphs or with a neural network that tolerates wear, dirt, and unusual fonts.
  • Post-processing applies syntax rules, such as which character patterns a given state uses, detects the issuing state, and attaches a confidence score. Vendor architecture overviews note that such scores and checks are often added by optional AI modules.

The trend is toward end-to-end deep-learning models that read the plate directly from the image. That removes hand-built stages, but it also makes the confidence score more important, because a single model can be confidently wrong in ways a segmented pipeline might have caught.

What limits accuracy: light, weather, angle, and speed

ALPR accuracy is not constant. A peer-reviewed study of ALPR performance on distorted images found that training models on low-illumination images improves detection, yet heavy fog or frost can still push detection accuracy to very low levels, even for advanced models.

Training helps, but weather can still win: extreme conditions such as heavy fog or frost can drop detection accuracy sharply, even for well-trained models. (source)

Common failure patterns show up across deployments:

  • Misreads of look-alike characters, such as 8 and B or 0 and O.
  • State misclassification, where the characters are right but the issuing state is wrong.
  • False positives and negatives from dirty, bent, or obstructed plates.

Deployments get better read rates by controlling what they can: a consistent mounting height, infrared illumination for night reads, a lens matched to the lane width, and regular lens cleaning.

Where ALPR is used

Different jobs ask different things of the same technology:

  • Law enforcement uses hot lists to flag stolen vehicles and wanted persons, and stored reads to build an investigative timeline of where a vehicle has been.
  • Tolling and parking use plate reads to bill drivers automatically and to measure how long a car stayed, replacing tickets and gates. Industry sources also point to ticketless parking and journey-time calculations.
  • Private security relies on it for gated communities, campuses, and lot monitoring, where the question is simply whether a plate is on the approved list.

The use case drives the design. A toll lane needs fixed, high-speed capture. A patrol car needs mobility. A gated facility can keep retention short because it only needs an access decision, while an investigative program may keep reads far longer.

Data handling and the patchwork of U.S. rules

Federal and state rules may affect how the technology is governed. State-level policies vary, and many require an audit trail for data access, a justification for each query, and a cap on retention, with some states limiting it to three years.

Good operational practice, in line with NIJ-style guidance, tends to include:

  • Audit logs showing who searched what, and why.
  • Restricted query rights tied to training.
  • Regular cleanup of hot lists so expired alerts do not linger.

A locally downloaded hot list can be stale the moment it is copied, while a live query reflects the current record. That trade-off between speed offline and freshness online is one every agency has to weigh.

Why a person confirms every hit

A plate read is a lead, not a verdict. Practitioner guidance states that operators must confirm the read is accurate, that the issuing state matches, and that the reason for the alert is still current before acting. Interfaces show the overview image next to the plate close-up so the check can be quick. Training and procedure exist to prevent a wrongful stop caused by a single misread character.

ALPR image review and alert verification flow

Our take: know what the camera can see

ALPR is genuinely useful for recovering stolen cars, running toll lanes, and securing private lots, and it also creates a searchable record of where ordinary people drive. Understanding the pipeline, from capture through OCR to matching, helps drivers and security teams make clear-eyed choices about privacy, retention, and risk. For a wider view of what is available, see our guide to license plate privacy options and U.S. rules.

— GoGhost

A privacy option for drivers who want one

If limiting automated plate capture on your own vehicle matters to you, we make the Ghost Curtain Cover, a blackout cover for your plate, priced at 149.99 USD one-off. It is offered in electric and mechanical versions. Check the product listing for fit, installation, and shipping details. It is for privacy, not for evading lawful enforcement, so check your local rules first and read our legal guidance linked above.

Goghost

FAQ

Is ALPR the same as ANPR and LPR?

Yes. These terms describe the same underlying technology, and “ALPR” is the common U.S. term while ANPR is used in other regions. LPR is simply the shortest form of the same name.

How does ALPR actually read a plate?

It captures an image, locates the plate, straightens and normalizes it, splits it into characters, and runs OCR. Syntax checks and a confidence score then validate the result before it is matched against a database.

Is number plate recognition AI?

Modern systems usually are. Object-detection and neural OCR models handle plate finding and character reading, and AI modules often add confidence scoring and context, though older systems relied on template matching.

Do officers act on an ALPR hit automatically?

No. Standard practice is for an operator to verify the read, the issuing state, and the current alert reason before taking enforcement action.

Sources