Glossary / Computer vision

What is automatic number plate recognition?

Automatic number plate recognition (ANPR) is computer vision that detects a vehicle's registration plate in a video frame, isolates it, and converts the characters to text. On Indian roads its accuracy depends far more on camera placement, shutter speed and plate condition than on the recognition model itself.

Reviewed by Dr. Vivek Gujar, Chief Strategy Officer  ·  Updated 31 August 2026  ·  7 min read

ANPR is also written as ALPR (automatic licence plate recognition) or simply LPR. The three terms describe the same capability. It is the most widely deployed video analytic in India because the output is unambiguous: a string of characters that either matches a record in your system or does not. Unlike behaviour analytics, there is no interpretation layer to argue about.

That clarity is also why ANPR projects fail in a specific, measurable way. When the read is wrong, everyone knows immediately, because a barrier does not lift or a wrong vehicle gets logged. This page covers what the pipeline actually does, what the camera has to deliver for it to work, and how to read an accuracy claim properly.

How the ANPR pipeline works

A production ANPR system runs five stages on every frame it decides to process. Understanding the sequence matters because failures are almost always attributable to one stage, and the fix is different for each.

1. Vehicle detection. An object detector finds vehicles in the frame and gives each one a bounding box. This narrows the search area and reduces false plate detections on signage, stickers and reflections.
2. Plate localisation. A second model finds the plate within the vehicle box. On Indian vehicles this is harder than in Europe because plate position varies widely, particularly on two-wheelers and commercial vehicles with bull bars or aftermarket mounts.
3. Rectification. The plate crop is deskewed and normalised. A plate photographed at an angle is a trapezoid, and the OCR stage expects a rectangle. Beyond roughly 30 degrees of horizontal or vertical incidence, rectification starts discarding information rather than recovering it.
4. Character recognition. OCR converts the rectified crop to a character string, usually with a per-character confidence score attached.
5. Format validation. The raw string is checked against Indian registration formats. A read of MH12AB1234 is plausible. A read of MH12AB123Z is not, because the last four characters must be numeric, so the system can correct or reject it. This stage recovers a meaningful share of otherwise failed reads and is often the cheapest accuracy improvement available.

Stages one through four are model work. Stage five is business logic, and it is where India-specific tuning earns its keep. The same is true of multi-frame voting, where the system reads the plate across several frames as the vehicle approaches and takes the consensus rather than trusting any single frame.

Read rate and accuracy are different numbers

Most ANPR accuracy claims are unusable because they do not say which number is being quoted. There are three, and they compound.

End-to-end performance is read rate multiplied by plate accuracy. Ask for both, and ask over what sample.

What the camera has to deliver

Nearly every ANPR deployment that underperforms was under-specified at the camera, not at the model. Three parameters govern the outcome.

Pixel density at the plate

OCR needs a minimum character height in pixels. Below roughly 16 px of character height, reads degrade sharply; 20 to 25 px is the reliable working range. You can derive the pixel density you need from the physical plate.

Two-wheelers need close to twice the pixel density of cars for the same read quality. On a mixed gate, spec for the two-wheeler or accept that two-wheeler reads will be materially worse than car reads. This single point explains most disappointing results at housing society and campus gates, where two-wheelers are the majority of traffic.

Shutter speed

Motion blur destroys plate reads before resolution ever becomes the limiting factor. A vehicle at 60 km/h travels 16.7 metres per second.

At 1/500 s the smear is wider than the strokes it is meant to resolve. Fast shutters starve the sensor of light, which is why dedicated ANPR cameras pair them with strong infrared illumination and why the retroreflective coating on HSRP plates matters so much. Reading an HSRP plate under IR is a different, easier problem than reading a painted plate under IR.

Angle and lane discipline

Keep both horizontal and vertical incidence under about 30 degrees. Every degree past that costs rectification quality. A camera mounted high to survey a wide forecourt and also asked to read plates will do neither job well. ANPR generally wants a dedicated, tightly framed camera per lane.

Where Indian conditions break ANPR

Models trained predominantly on European or North American plates underperform here for reasons that have nothing to do with model quality:

• Plate heterogeneity. HSRP plates, older painted plates, hand-painted plates, decorative fonts and regional script variants all coexist on the same road.
• Two-wheeler share. Smaller plates, often two-line layouts, frequently mounted at an angle or partly obscured by a mudguard.
• Occlusion and contamination. Dust, monsoon spray, mud, bull bars, tow hooks and number plate frames.
• Two-line plates. Common on commercial vehicles and two-wheelers, and a distinct localisation problem from single-line plates.
• Lighting extremes. Direct low-angle sun on a reflective plate saturates the sensor as badly as no light at all.

The practical consequence is that a benchmark run on a public international plate dataset tells you very little about how a system will perform at your gate. Ask for footage from a deployment resembling yours, or run a two-week pilot on your own cameras before committing.

Running ANPR at the edge

ANPR is a strong candidate for edge processing because the decision is time-critical and the data is sensitive. A barrier has to lift while the driver is still at it, which makes a round trip to a cloud region a poor fit for the control loop. Processing on site also means the video never leaves the premises: only the plate string and a timestamp need to be transmitted.

In retrofit deployments, an on-site appliance pulls RTSP streams from the existing NVR, so the cameras and recorder already installed continue to do their job while the analytics layer is added alongside them. Where the cameras support ONVIF, discovery and stream configuration are considerably faster.

Cloud ANPR is the better choice in three situations, and it is worth being clear about them. If plate reads are logged for later analysis rather than driving a barrier, latency does not matter. If you are running a handful of cameras across many small sites, a single cloud tenancy is cheaper to operate than an appliance per site. And if you already run a cloud VMS that your team knows, adding a second on-premise system creates operational cost that the technical benefit may not repay. Edge earns its place when the decision is real-time, the camera count per site is high, or the data cannot leave the premises.

ANPR and DPDP

Practical implications worth settling before deployment rather than after:

• Define and document a retention period for plate logs and associated images, and enforce it automatically. Indefinite retention is difficult to justify against any stated purpose.
• Post notice at the point of collection. The gate is where the notice belongs.
• Keep matching against blocklists or resident registers on site where practical, so the linking of plate to individual happens within your own perimeter.
• Note that DPDP Rule 4 commences on 13 November 2026, with the Section 33 penalty provisions following on 13 May 2027.

This is general information rather than legal advice. Deployments involving resident or employee data should be reviewed by your own counsel.

Frequently asked questions