Published 10 November 2026 · 15 min read · Verification

Every plate camera is sold with two numbers, and buyers treat both as promises about the same thing. They are not. Capture asks whether the plate was recorded at all. Recognition asks whether the characters on a recorded plate were read correctly. A system can record every plate that passes and still misread one in fifty, or read every plate it sees perfectly and miss one in a thousand. Until the two are counted separately, a site that "misses plates" is a complaint with no diagnosis attached.

This page is the measurement half of the problem. How to define a miss so the number means something, how many passes a test actually needs before it can support a claim, how to stratify by condition so an average does not hide a failure, and how to locate the cause of a single miss in about five minutes. Our ANPR range is rated at a capture rate of 99.9 percent or better and a recognition rate of 98 percent or better. Neither figure survives contact with a real site unless somebody measures it the right way round.

Capture and recognition fail in different ways

The two failure modes look identical from the outside, because in both cases the operator says the plate did not come through. Inside the system they are unrelated events with unrelated causes and unrelated fixes.

A capture failure produces no record. The vehicle passes, the barrier does not lift or the log stays empty, and there is nothing to audit because nothing was written. A recognition failure produces a record with the wrong characters, or a record with a plate image and no characters at all. That second case is the dangerous one: the system appears to be working, the log fills up, and the errors only surface when someone reconciles the access log against the actual vehicle list a week later. By then nobody can tell which entry was wrong without going back to the images.

The reason to keep the two numbers apart is arithmetic. A site reporting 98 percent overall could be running at 100 percent capture with 98 percent recognition, which is a healthy system working exactly as specified. It could equally be running at 90 percent capture with 99 percent recognition, which is a badly aimed camera with a good engine behind it. The first needs nothing. The second needs a pole moved. Averaged into a single figure, both look the same, and the fix for one is useless against the other.

So the first thing to change on any acceptance sheet is the shape of the sheet. Two columns, two pass criteria, two owners. Capture belongs to whoever chose the mounting height, the trigger point and the shutter. Recognition belongs to whoever configured the character engine. Collapsing them into one percentage guarantees that the wrong team gets the support ticket.

What counts as a miss

Before counting anything, agree on what a miss is. Four outcomes are possible for every vehicle that crosses the read window, and only the first two should count as a successful capture.

The distinction that matters in practice is between a record that exists and a record that is usable. A triggered frame showing a smeared plate is not a captured plate, even though the log has a row for it. If your recorder writes a row on every trigger regardless of quality, the capture rate you calculate from that log will be inflated by exactly the number of unreadable frames, which is often the whole problem.

That is why the record schema matters more than the camera model. Each record needs to carry enough fields to be classified after the fact: a trigger timestamp, a full-scene frame, a plate crop, a confidence score and a reason code for rejected frames. Without a reason code, unreadable frames and untriggered passes both land in the same bucket, and you cannot tell a camera that missed the vehicle from a camera that saw it and produced rubbish.

ObservationWhat it meansWhere to fix it
Plate in the log, characters correctCapture and recognition both succeededNothing. Count it as one pass and one hit.
Plate in the log, characters wrong or blankCapture succeeded, recognition failedCharacter engine, plate crop quality, character set. Not an aiming problem.
Record exists, plate unreadable by eyeCapture failed even though a frame was writtenShutter, focus, glare, occlusion. Fix the image, not the engine.
No record at allCapture failed at the triggerVirtual loop position, radar alignment, loop detector, trigger threshold.

Four possible outcomes for one vehicle pass

How many passes before the number means anything

Here is the uncomfortable part. A capture rate is a probability, and a probability cannot be proven by a short test. The standard shortcut is the rule of three: if a test produces zero failures across n passes, the 95 percent confidence upper bound on the failure rate is 3 divided by n.

Put numbers in it. One hundred clean passes put the failure bound at 3 percent, so the test supports a capture rate of 97 percent or better and nothing more. Three hundred passes support 99 percent. One thousand passes support 99.7 percent. To support 99.9 percent you need three thousand consecutive passes with zero failures, under conditions you can defend as representative.

That is where most acceptance arguments actually come from, and it is worth saying plainly: three thousand passes is a day and a half of continuous traffic, and no integrator signs off that way. The 99.9 percent figure is a design and production figure, established under controlled conditions across a large sample. It is not something a site visit can confirm, and asking a commissioning engineer to demonstrate it is asking for a number nobody can honestly produce.

The practical substitute is a stratified zero-failure test. Break the site into the conditions it actually sees, run fifty passes in each, and require zero failures in every stratum. Fifty clean passes only prove 94 percent per stratum, but fifty clean passes in six different conditions is a far stronger statement about the installation than one thousand passes in perfect midday light. You are testing for the absence of a systematic fault, which is what a commissioning test can actually establish.

Stratify by condition, never pool

An average is the enemy of a diagnosis. A site reporting 99.5 percent overall can be hiding a night-time backlit stratum running at 85 percent, and the arithmetic will never show it. Split the test before you run it, not after.

The strata below cover the conditions that change the optics rather than the ones that change the traffic volume. Each has a distinct failure signature, which is the point: if a stratum fails, the signature tells you which setting to touch before you climb the pole.

Set the pass count per stratum from how much exposure that condition gets in service, with a floor of thirty. Do not pool the results. Report them as a list, because a list of six numbers with one weak entry is a report an operator can act on, and a single average is not.

ConditionWhat it stressesPassesFailure signature
Daylight, dry, frontal approachBaseline. Should be perfect if anything is.50Failures here mean geometry or focus, not lighting.
Night, oncoming headlightsExposure range against a very bright source.50Plate blown out, characters merged into white.
Night, no vehicle lightingFill light reach and evenness.50Plate too dark, low contrast at the far edge of the window.
Low sun, backlitDynamic range and silhouette.30Vehicle becomes a black shape, plate unreadable.
High approach speedShutter and trigger timing.50Directional smear across the characters.
Reflective or soiled plateCharacter separation under glare or dirt.30Partial reads, confident but wrong characters.
Trailer or tow bar, second plateMultiple plates in one frame.30Right image, wrong plate selected.

Minimum strata for a site acceptance test

Diagnosing a miss in five minutes

Most misses are diagnosable from the log alone, in the order below. Work the steps in sequence and stop at the first one that fires; the later steps are wasted effort if an earlier one already explains the record.

Step one, is there a record at all? If the vehicle produced no row, the failure is upstream of the image. Check the virtual loop against the frame: a loop drawn where the plate is small, or drawn partially outside the lane, will miss the vehicle entirely. On radar-triggered sites, confirm the beam covers the lane and is not clipping the adjacent one. On inductive loops, a failed loop reads as an empty lane, not as an error, so verification needs a physical test with a vehicle.

Step two, is there a frame but no readable plate? Open the image and look at it as a human before looking at the character output. Blur that runs in one direction is shutter speed. Uniform blur across the whole frame is focus. A white rectangle where the plate should be is exposure against the fill lights or the headlights. A missing plate is occlusion, and the culprit is usually an accessory on the vehicle, not the camera.

Step three, does frame rate give you enough chances? At 25 frames per second the camera produces a frame every 40 milliseconds. A vehicle at 150 km/h covers 41.7 metres per second, which is 1.67 metres between frames. Across a 35 metre usable window that is roughly twenty-one frames, and the vehicle is only inside the sharpest part of the window for a fraction of those. Move the trigger point two metres and you have spent one or two of your chances. This is why trigger placement matters more at speed than any other single setting, and why a site that works at 30 km/h can fail at 90 with nothing else changed.

Step four, is it actually a recognition problem? If the image is sharp, correctly exposed and the plate is plainly legible to a human, the capture stage did its job and the character engine is the subject. Look at the crop, not the scene: a plate that sits half outside the crop region, or a character set the engine was not configured for, produces confident wrong output rather than a low score.

One counter-intuitive case worth naming. Adding fill light to a reflective plate can make capture worse, not better. Retro-reflective plate film is designed to send light straight back to its source, so a close, powerful illuminator washing the plate at a shallow angle can return a saturated block of white where the characters were. If night capture is worse on the reflective plates than on the dirty ones, the answer is usually less light or a different angle, not more light.

The plate is a variable you do not control

Part of the capture rate is decided before the camera is chosen, by the plates themselves. A camera is specified against plates that meet the legal standard at that location, and every deviation from that standard costs something.

Soiling is the common one. A plate carrying a film of road spray still has edges, and the engine often reads it, which is why dirty plates fail in ones and twos rather than in blocks. Retro-reflective film is the harder case, as described above: it is brightest exactly when you least want it to be. Frames, brackets and dealer surrounds eat character edge margin. A tow bar, a bike rack or a trailer pushes a second plate into the frame, and now the question is not whether a plate was captured but which of two plates the system chose to keep.

Temporary and paper plates are a separate category. They fade, they curl, and they are often mounted where a permanent plate would not be. Do not fold them into the main figure and then argue about the camera; report them as a separate line so the number stays honest.

Character set is the last variable and the one most often ignored in international deployments. A system installed across more than a hundred countries has to handle more than Latin characters, and non-Latin plates do not always define character height the same way. Confirm the engine has the character set for the market before commissioning, because no amount of aiming will recover a script the engine was never trained on.

Capture evidence, not just a count

A capture rate is a statistic. What the operator actually needs is evidence, and evidence has requirements that a count does not.

A usable record carries four things: a full-scene frame that shows which vehicle in which lane, a plate crop sharp enough to read by eye, a timestamp that survives timezone conversion, and a direction. Drop any one and the record cannot settle a dispute. The scene frame without the crop cannot prove the characters. The crop without the scene frame cannot prove which vehicle it came from. Both without direction cannot tell an entry from an exit at a two-way lane.

Retention is arithmetic, not a specification. Days of storage equals capacity divided by bitrate times 86400 seconds. Work the bitrate out from the actual stream rather than the datasheet, because H.265 at 1080p25 varies by several times between a quiet yard and a busy gantry. Do that sum before promising a retention period, and remember that plate records are small: the images are what fill the disk, not the text.

Where the architecture sits also affects how much can fail. The recognition engine runs on board at 1 TOPS, so a site does not need a server in a cabinet to turn frames into plate strings. That removes a whole class of failure that shows up as "the camera is fine but nothing reaches the platform", which is usually the server or the link rather than the camera. Streams leave over ONVIF Profile S into a recorder, and the 4G models can carry a 512 GB NVR internally for sites with no server room. Power over PoE or 12 V DC keeps the camera on the same supply as the switch, which is the difference between a site that survives a short outage and one that needs a visit.

Related pages

Installation questions we get asked

What counts as a good license plate capture rate?

Our ANPR range is rated at a capture rate of 99.9 percent or better and a recognition rate of 98 percent or better, measured under controlled conditions over a large sample. On a live site the honest target is different: zero failures within each condition the site actually sees, tested separately for daylight, night with headlights, night without, low sun, high speed and reflective or soiled plates. A site that passes every stratum is in good shape; a site that reports one blended average is not telling you anything.

How many vehicle passes do I need to test a capture rate?

Use the rule of three: with zero failures across n passes, the 95 percent upper bound on the failure rate is 3 divided by n. One hundred passes support 97 percent, three hundred support 99 percent, one thousand support 99.7 percent, and proving 99.9 percent needs three thousand consecutive clean passes. That is why site acceptance is normally done as a stratified zero-failure test with thirty to fifty passes per condition rather than as an attempt to reproduce the datasheet figure.

Why does capture get worse at night when the camera has fill lights?

Usually because of the plate, not the light. Retro-reflective plate film returns light directly to its source, so a close illuminator at a shallow angle can saturate the plate into a white block and erase the characters. Headlights aimed at the camera cause the same effect from the other direction. If the reflective plates fail while the dirty ones read fine, reduce the fill light or change its angle rather than adding more.

Is a missed plate always a capture problem?

No, and separating the two is the first diagnostic step. If the log has a record with the wrong characters, capture succeeded and the character engine failed. If the log has a frame but the plate is unreadable to a human, capture failed on image quality. If there is no record at all, capture failed at the trigger. Each of those three has a different fix, and treating a recognition problem as an aiming problem wastes a site visit.

Does frame rate limit capture at high speed?

It sets the number of chances you get. At 25 frames per second a frame arrives every 40 milliseconds, and a vehicle at 150 km/h covers 41.7 metres per second, which is 1.67 metres between frames. Across a 35 metre usable read window that is about twenty-one frames in total, with only part of them in the sharpest region. Trigger placement therefore matters more at speed than any other setting.

Can a poor capture rate be fixed after installation?

Often yes, without moving the pole. Move the trigger point so the plate is largest when the frame is taken, set the shutter for the approach speed, recheck focus after zooming, and examine whether the fill light is helping or hurting on reflective plates. If all four are already right and one stratum still fails, the remaining cause is usually geometry, and that is a pole or lens decision rather than a settings one.

Send us the misses and we will tell you where they start

If a site is missing plates, send us the log extract and a handful of the frames that failed, plus the mounting height, the working distance and the approach speed. With those four numbers we can usually tell you whether it is a capture problem or a recognition problem before anyone climbs the pole.

Request a capture diagnosis

Or email goingtech@sz-going.cn · +86 755 8395 5458