Defining accuracy in ANPR systems

A practical method for testing plate detection, exact reads and the evidence delivered to traffic engineers

Technical whitepaper · Version 1.1 · 7 November 2026

The denominator is part of the result.

Download the whitepaper (PDF) · 14 pages · Word document

Inside this paper

Accuracy starts with the question
Follow the denominator through the chain
Define the Australian plate population
Design an Australian capture envelope
Freeze motion and preserve character contrast
Match the test to the delivered evidence
Size the sample for the acceptance decision
Accept the configured site
A reusable Australian field-test matrix
Locate the failure, then test the correction
XVision field result: Casey Complex
Sources

This paper defines a repeatable method for evaluating automatic number plate recognition (ANPR), also called automatic licence plate recognition (ALPR) or LPR. It covers capture, detection, recognition and the delivery of usable records.

The numeric example is illustrative. The Casey Complex result is a separate field result. Source references identify the basis and context of each attributed statement.

Accuracy starts with the question

Traffic and transport engineers need to know how far they can trust plate-derived evidence before using it in a parking study, journey-time assessment or road investigation. A headline accuracy percentage leaves that question unanswered unless it states what succeeded, what was counted and where measurement ended.

Finding a plate is different from reading its complete registration. Reading an already-detected plate is different from delivering the correct registration for every eligible vehicle passage. A correct engine string lost before the agreed output is an engine success and a whole-system failure.

Measure the chain, not one favourable stage

Use one passage as the counting unit. Declare its target plate, lane, direction, capture zone and reporting boundary. Build the reference independently of the system, then report detection, exact reads, wrong reads and no-reads together. Retain the configuration and conditions beside every result.

The primary outcome is exact reads per eligible passage. Conditional measures diagnose losses. Audit false detections, duplicates and delivery problems separately.

Keep three kinds of evidence separate

Evidence class	Role in this paper
Published engine guidance	Plate Recognizer, the engine supplier, provides capture and processing reference material [V1-V4]. Its recommendations are starting points for site testing.
XVision evaluation framework	The passage definitions, reference method, test matrix and acceptance process specify how to evaluate evidence.
Measured field result	Casey Complex is reported as plate detection for its stated deployment and period [X1].

Engine-vendor demonstrations do not establish XVision performance. Design targets become measured results only after a defined test. The illustrative example in Section 02 teaches the calculations; it is not a deployment result.

Scope and limits

This is a technical evaluation framework, not legal advice or a catalogue of software functions. Set thresholds from the intended engineering task, operating conditions and consequences of error. Engineers retain responsibility for interpretation and acceptance. A result applies to the population, configuration and conditions actually tested.

How XVision uses this framework

Agree the engineering question, capture envelope and test matrix. Retain evidence sources and limitations. Record the acceptance decision and follow-up checks. Confirm available records and the supported data path before agreeing the evaluation scope.

Follow the denominator through the chain

For one declared target plate per eligible passage: N is the independent passage census; D is passages with a correctly localised target plate; R is passages with an accepted text result at the chosen output; E is passages with an exact, correctly associated full-plate result. Thus E is within R, R within D and D within N.

Figure 2.1. Conceptual hierarchy. Unmatched false detections and duplicate records are audited separately.

Illustrative numeric example

Assume 1,000 independent passages, each with one six-character plate and independently known truth. All results use the same configuration and output boundary: N = 1,000; D = 960; R = 940; E = 900. Of 60 no-reads, 40 are missed detections and 20 are detected plates without accepted text. The 40 wrong accepted strings each contain one character substitution. There are no insertions, other deletions, duplicates or unmatched records in this example.

Metric	Calculation	Result
Detection rate	D/N = 960/1,000	96.0%
Exact reads, end-to-end	E/N = 900/1,000	90.0%
Exact reads, conditional on detection	E/D = 900/960	93.75%
Exact-read precision among accepted reads	E/R = 900/940	95.74%
Wrong reads per passage	(R - E)/N = 40/1,000	4.0%
No-reads per passage	(N - R)/N = 60/1,000	6.0%
Character accuracy, all passages: 1 - CER	1 - (40 + 360)/6,000	93.33%
Character accuracy, accepted strings only	1 - 40/5,640	99.29%

The 360 deleted characters are the six missing characters in each no-read. Character accuracy among accepted strings looks strong while end-to-end exact reads remain 90.0%. The exact, wrong and no-read shares sum to 100%. Wrong reads among accepted strings are 40/940 = 4.26%.

The independently calculated Wilson 95% interval for 900/1,000 is 87.98%-91.71%; inputs and method appear in Section 07. For the same nested population, E/N = (D/N) × (E/D). Multiplying rates from different studies does not satisfy this identity.

Define the Australian plate population

Australian plates have state, territory and series-specific dimensions and character arrangements. A single six-character grammar or an overseas plate width cannot represent the whole operating population. Use the smallest and most demanding in-scope formats when designing capture.

Official Australian reference	Verified detail	Evaluation consequence
WA state-series plates [A1]	Motor vehicle: 372 × 134 mm. Motorcycle: 254 × 100 mm full-size or 190 × 100 mm small.	Test smaller motorcycle targets separately; do not assume car-plate pixel density.
WA optional-plate form [A2, p.1]	Normal: 372 × 134 mm; slimline: 372 × 100 mm; half-size: 190 × 134 mm; three-quarter: 315 × 134 mm.	Width and height both vary. Measure characters as well as the plate outline.
WA optional formats [A2, pp.2-3]	Some custom motor-vehicle formats allow up to nine characters. Other series prescribe letter/number groups; spacing rules vary.	Preserve variable length, letter/digit positions and meaningful spaces in the reference.
Queensland series announcement [A3]	Six-character pattern introduced as three digits, two letters and a digit, such as 000-AA0. It also describes later position changes.	Do not enforce a three-letter/three-digit rule or merge zero with letter O.

These are sourced examples, not a national size catalogue or proof that a format has been observed at the site. Include plates from all required jurisdictions, older series, personalised plates, slimline plates, motorcycles, trailers, heavy vehicles and auxiliary plates in the inventory. Record actual dimensions, text length, layout, colour/reflectivity and expected frequency. Select front or rear capture to suit the required vehicle classes and plate presentation.

Exact-match rules

An exact read matches every required registration character in the agreed order. One substitution, missing character or extra character fails the full-plate result. Agree case, decorative separators and space normalisation before scoring; retain both raw and normalised text. Keep letter O and digit 0 distinct. Define reading order for stacked or multi-line layouts. Count only the selected automatic result, not an alternative list containing the answer.

For trailers or vehicles presenting multiple plates, define the target and its passage association. Report jurisdiction correctness separately when the same text in different jurisdictions would identify different vehicles.

Character accuracy is a different measure

Character error rate, CER = (substitutions + deletions + insertions) / reference characters, uses a declared edit alignment. This paper labels 1 - CER as character accuracy; CER can exceed one when insertions are numerous. State the population. Missing strings contribute all their characters as deletions in an all-passage calculation. Score full plates directly rather than assuming independent character errors.

Design an Australian capture envelope

Record an operating envelope for each lane and direction: target plate formats and dimensions, capture range, near/far lane positions, approach/turning paths, measured speeds in km/h, usable observation time and day/night conditions. Define the required bounds from the actual road and task, then test their demanding combinations.

Capture quantity	Definition at the plate	Field measurement
Yaw	Horizontal obliqueness between viewing direction and plate normal	Range across the zone and turning paths
Pitch	Vertical obliqueness between viewing direction and plate normal	Mount height, road grade and plate tilt
Roll	Plate-baseline rotation in the image	Device orientation, crossfall and vehicle lean
Pixels on plate	Projected plate width in the actual engine input	Smallest target, far lane, crop and resizing
Character detail	Character height, stroke width and separation in that input	Small fonts, long strings, stacked text and compression

Figure 4.1. Capture-geometry reference table. The rotations describe different effects and are measured at the target, not only against the road centreline.

Plate Recognizer recommends roughly 100 pixels across the plate and horizontal/vertical angles no greater than 45 degrees in its setup guide [V1]. Its resolution guide describes about 100 pixels as a practical target, with 120-150 pixels adding margin in difficult scenes [V2]. These are engine-supplier recommendations, not XVision pass limits. Test combined yaw and pitch rather than treating each maximum as independent permission to use both together.

Illustrative Australian pixel calculation

For an approximately front-facing plate, estimate:

$P = \frac{H W_p}{W_s}$

Here P is plate width in pixels, H is input image width in pixels, Wp is physical plate width and Ws is scene width at the plate plane, in matching length units [V2].

With H = 1,920 and an illustrative target P = 100, a 372 mm WA motor-vehicle plate gives Ws = 1,920 × 0.372 / 100 = 7.14 m. A 190 mm WA small motorcycle plate gives 3.65 m [A1]. The same broad view therefore provides very different pixel coverage for these two Australian targets.

This is a projection estimate for a front-facing plate, not an oblique-view guarantee. Confirm measurements in real engine-input frames. Select lens, field of view and resolution together; digital enlargement cannot restore lost detail. Set character-detail targets through reads of the smallest required characters rather than inventing a universal pixels-per-character threshold.

Freeze motion and preserve character contrast

Longer range reduces projected plate size unless optics or input resolution change. Longer focal lengths narrow the view and increase sensitivity to vibration and focus errors. Check near/far focus, mounting stability, edge detail and time in the usable zone. Plate Recognizer advises minimising range and gives 35 metres as its maximum advisable distance [V1]; use that as supplier setup guidance, not a warranted XVision range.

Shutter duration and frame rate do different jobs

A first-order diagnostic is:

$b \approx u t$

Here b is motion smear in pixels, u is measured image-plane plate speed in pixels per second, and t is exposure in seconds. Measure u from actual site frames; road speed alone omits viewing geometry. Establish a blur budget with representative Australian plates at the highest in-scope speed and most demanding trajectory. Shorten exposure to meet that budget, then provide enough light and control gain to retain character contrast.

A higher frame rate gives more opportunities; it does not sharpen a blurred frame. Record the actual analysed frame rate, compression, resizing, dropped frames and processing load. Plate Recognizer distinguishes effective frame rate from capture frame rate and explains prediction-delay trade-offs [V4]. Confirm that the shortest usable passage offers clear analysed frames and that the result arrives by the agreed deadline.

Condition	Failure mechanism	Required check
Low light and mode switching	Noise, slow exposure or focus shift	Moving plates at operating speed; exposure, gain and day/night focus
Sun, headlights and wet glare	Saturation or flare hides strokes	Demanding sun directions, approach paths and reflections
Infrared illumination	Weak or uneven return; reflective plate saturation	Beam coverage, near/far return, wavelength compatibility and focus
Rain, spray, fog and dust	Droplets, veiling and contrast loss	Observed severity, optical-window condition and plate detail
Occlusion and queues	Target partly or wholly hidden	Following gaps, truck traffic and adjacent lanes
Dirty, bent or damaged plates	Missing strokes and changed geometry	Independent truth and defect labels retained in scoring

Plate Recognizer's capture guidance covers exposure and IR [V1-V2]; its difficult-image page demonstrates selected challenges [V3]. IR does not remove glare or weather effects. Test the imager, illumination and local plate materials together. Include interacting conditions, such as far-lane night capture in rain. A digitally darkened still image does not replace a moving-vehicle night test.

Match the test to the delivered evidence

Boundary	Included	Boundary record
Generic camera capture	Optics, imaging sensor, exposure and illumination	Recorded plate detail at operational settings
Recognition engine	Detection and decoding of specified input	Input file and raw engine output
Configured XVision unit	Agreed sensing and processing up to its declared output	Device configuration and available event record
Whole system	Capture, recognition, filtering, association and final delivery	Correct records at the agreed evidence output

The evaluation plan specifies available stage records; this table does not promise a particular implementation. Compare independent passages with source capture, engine events and delivered records wherever those records exist. Keep scheduled service availability separate from recognition during available periods, then report the whole-service outcome where the reference permits it. A missing service interval is not zero traffic.

Build ground truth without seeing system answers

Select observation blocks randomly or by a predeclared systematic schedule across representative days. Cover lanes, directions, peak/off-peak traffic, speeds, plate formats, lighting and weather. Oversample rare critical conditions, report them separately, and use independently justified population weights for any combined estimate.

Reviewers log each passage, target plate, time, lane, direction and condition without seeing predictions. Use authorised independent evidence when the operational image cannot establish the string. Double-review a predeclared subset and disputed cases, adjudicate differences and retain corrections. Keep tuning and acceptance periods separate, including repeated visits and adjacent frames.

Fix clock alignment, time zone, matching window and trajectory rules before comparison. Match with independent time/path evidence rather than predicted plate text. Retain unmatched outputs and ambiguous associations for the record audit.

Preserve the denominator when truth is difficult

N comes from the independent census, including in-scope misses and occlusion. Record vehicles without a target plate as a distinct category. An all-vehicle service rate retains them in N; a plate-present rate uses a separate declared population. D/N is detection yield per passage, not localisation recall when N includes absent plates.

When independent evidence cannot establish plate text, retain the passage in the census and label text truth unscorable. Report exact-read accuracy on the reference-known subset together with its coverage of N. Bound an all-passage estimate by treating unknown cases first as failures, then successes. Use authorised access, retention and handling rules; publish aggregates rather than identifying plate imagery.

Size the sample for the acceptance decision

Choose the metric, critical strata, confidence level, precision and acceptance rule first. Count independent passages, not repeated frames. Conditional exact-read accuracy needs enough detected plates D; planning from N alone understates that requirement. Allow for unusable truth records, repeated vehicles, platoons and site/time effects.

Illustrative numeric example: sample planning

A first planning approximation for a proportion is:

$n \approx \frac{z^2 p(1-p)}{h^2}$

Using p = 0.5, desired half-width h = 0.03 and z = 1.96 for an approximate two-sided 95% interval gives n = 1,067.111..., rounded up to 1,068 independent passages. This is a planning illustration, not a universal test size or an acceptance threshold. Confirm final sample size against the intended interval across plausible outcomes. Plan each critical stratum separately. For a minimum-performance decision, also plan statistical power at a defensible alternative and the required lower confidence bound.

Wilson interval: inputs and calculation

NIST describes the Wilson interval for binomial proportions [N1]. For x successes in n independent trials, let p-hat = x/n. Calculate centre c and half-width w:

$c = \frac{\hat p + z^2/(2n)}{1+z^2/n}$

$w = \frac{z\sqrt{\hat p(1-\hat p)/n + z^2/(4n^2)}}{1+z^2/n}$

The interval is [c - w, c + w]. For Section 02: x = 900; n = 1,000; p-hat = 0.9; z = 1.96. The calculation gives c = 0.89846924 and w = 0.01862160. Limits are 0.87984764 and 0.91709084, or 87.98%-91.71%. The selected 95% confidence level is a reporting convention, not an accuracy target.

Report successes, denominator, interval method, confidence level and population with every rate. Use the agreed one-sided bound for one-sided acceptance. Zero observed errors still requires an upper error bound.

Correlation and stopping rules

Use a survey-aware analysis or resample whole independent blocks/sites when observations are clustered, preserving strata and weights. Declare block selection and check sensitivity. An unweighted binomial interval does not describe a weighted overall result. Predeclare stopping rules and multiple acceptance criteria; repeated testing until a pass overstates certainty. Intervals quantify sampling uncertainty, not biased truth or untested conditions.

Accept the configured site

Step	Commissioning action	Evidence retained
01	Agree the task, target population, boundary, metrics and pass rules	Test plan and completed matrix
02	Survey the Australian capture envelope	Plate inventory, speeds, lens, range, angles and pixel measurements
03	Freeze settings and versions	Device/firmware/engine versions, exposure, regions, zones, thresholds and event rules
04	Collect independent reference blocks	Census, blind labels, truth coverage and adjudication
05	Match and score each declared boundary	D, R, E, wrong reads, no-reads, false detections, duplicates and output losses
06	Apply the acceptance rule	Per-stratum results, intervals, limitations and engineering decision
07	Correct and retest where required	Original failure, change record and fresh held-out test

Set end-to-end exact-read and wrong-read criteria from the engineering task and error consequences. Specify whether each critical condition must pass and whether a confidence bound governs the decision. Include availability, timely delivery, duplicate handling and completeness where they affect the task. Strong daytime results do not offset a failed required night stratum.

Confidence scores are not measured accuracy

An engine score expresses model certainty about a prediction. It does not establish site accuracy or prove an individual registration. Select thresholds on development data, then freeze them for acceptance. On held-out data, compare score bands with observed exact-read proportions and plot accepted coverage against wrong-read rate. Increasing the threshold rejects some correct reads as well as errors; it cannot recover missed plates.

Plate Recognizer's Stream Health Score concerns processing performance on host hardware [V4]. Use it as an operational diagnostic, not an accuracy percentage. Recheck score calibration after engine or scene changes.

Audit output records as well as passages

Wrong accepted text differs from no output. For known-truth passages, report (R - E)/N and (N - R)/N, plus wrong reads among accepted strings, (R - E)/R. An empty denominator is not estimable. For all accepted output records, separately report the share that is exact and uniquely matched to a real target passage. Include unmatched false detections and duplicates in that audit; E/R in Section 02 excludes those extra records.

Define the permitted prediction delay and final-output deadline. Report timely correct records separately from eventually correct records. Test both under peak load. A cropped-image replay diagnoses an engine; acceptance of the installed system requires the agreed field evidence.

A reusable Australian field-test matrix

Complete the levels, sample design and acceptance criteria before testing. Split a row wherever a lane, direction or condition requires its own acceptance decision.

Stratum	Scope to declare	Capture / condition record	Scoring evidence
Baseline day	Every lane/direction; speed bands in km/h	Near/far range, angles, plate and character pixels	N, D, R, E; rates and intervals
Night	Required speeds and vehicle classes	IR, focus, exposure, gain and glare	Separate night result
Dawn / dusk	Sun-facing and shaded paths	Sun direction, flare and mode changes	Wrong reads and no-reads
Wet / reduced visibility	Required traffic mix	Rain, spray, fog/dust and optical-window condition	Observed severity; stratum result
Geometry limits	Far lane, turns, grades and zone edges	Worst yaw/pitch/roll; smallest plate detail	Per-trajectory results
Congestion	Queues, close following and truck traffic	Occlusion and usable observation time	Census includes hidden targets; duplicates
Australian plate diversity	State/territory series, personalised, slimline, motorcycle, trailer, heavy vehicle and auxiliary	Dimensions, length, layout, reflectivity, dirt/damage	Independent text truth; exact text and jurisdiction correctness
Load and recovery	Peak demand, restart and missing service	Analysed frames, gaps and output timing	Engine versus delivered record; availability

Result-sheet template

Field	Complete for this test
Identity and scope	Test ID; site/unit; dates/time zone; lane/direction; target plate; reporting boundary
Capture and configuration	Plate inventory; range/angles; lens; speed bands; pixels; exposure/IR; versions/settings
Reference and selection	Block selection; truth source/review; eligibility; planned sample; weights; correlation
Counts	Scheduled/observed time; N; absent targets; known/unscorable truth; D; R; E
Errors and uncertainty	Wrong/no-reads; unmatched detections; duplicates; association/output loss; interval method/limits
Decision and traceability	Criterion; result; evidence file IDs; reviewer; change/retest ID; next review trigger

Retain raw and normalised text, original outputs and corrections in the authorised evaluation record. The published result should state the numerator, denominator, conditions, uncertainty and decision without exposing identifying evidence.

Locate the failure, then test the correction

The examples below are illustrative mechanisms, not XVision trial observations. The plate string is synthetic.

Failure example	Metric effect	Next check
ABC123 accepted as A8C123	Detection success; wrong exact read	Character detail, saturation and score bands
Plate found but text rejected	D increases; R does not	Threshold coverage and rejection reasons
Far-lane motorcycle plate too small	Miss or misread	Actual engine-input pixels and lens/FOV
Night plate saturated white	No-read or confident wrong read	Exposure, IR return and moving-vehicle contrast
Truck hides the target	In-scope end-to-end failure	Occlusion, gaps and capture-zone coverage
Sign becomes a plate record	Unmatched false detection	Nuisance objects and detection zones
One passage produces repeated events	Duplicate records	Association and event-separation settings
Correct engine read never arrives	Whole-system failure	Stage logs, timing and service gaps
Results swapped between close vehicles	Association failure	Independent trajectories and clock alignment

Retain the original failure, reference, condition and configuration. Tag causes after scoring; label live, replay and synthetic evidence separately. Exposure, lens, zone or threshold changes create a new configuration; confirm the correction on fresh evidence rather than deleting the original error.

Keep the accepted baseline visible

Set review frequency from operational risk, seasonal exposure and the cost of wrong records. Select rolling reference samples without inspecting successful reads first. Track end-to-end exact reads, wrong reads, no-reads, duplicates and truth coverage by lane and condition, using the commissioning definitions.

Repeat the relevant strata after changes to pose, lens, exposure, illumination, regions, thresholds, engine/firmware or event logic. Compare configurations on held-out reference material and verify the installed result. Preserve the last accepted baseline and every change record.

Detect deterioration before it changes the study

Inspect mounting movement, dirty optical windows, vegetation, new lighting, roadworks and changing traffic paths. Investigate falling plate pixels, rising rejected reads, unexpected formats, duplicate events and missing intervals. Record the owner, trigger and action for each check. A healthy processing score does not replace a reviewed sample of actual outcomes.

XVision field result: Casey Complex

The public Casey Complex case study reports the following result [X1]:

Validated over 26 April to 24 June 2023 across 18,310 dwell-time records from the three plate-reading units, achieving a 96.8% plate detection rate.

This is plate detection, not exact full-plate reading. It applies to that deployment and period. The quoted total counts dwell-time records. No passage denominator, plate count or exact-read rate is derived from it. Another site requires its own acceptance test.

Practical evaluator checklist

A useful result statement names the site and period, population and truth coverage, configuration, system boundary, metric formula, numerator/denominator, interval, conditions, limitations and engineering decision. It keeps supplier guidance, evaluation targets and measured outcomes distinct.

Discuss a Location

Contact XVision to discuss your location, capture conditions and ANPR evaluation requirements.

Discuss a Location

Sources

All sources accessed 7 November 2026. The Australian authority sources describe their stated jurisdictions and series. Engine-supplier pages are reference guidance, not measured XVision performance. Installed engine versions belong in the site test record.

ID	Source and context	Precise URL
A1	Transport WA, State series plates. Dimensions and target classes; live page.	https://www.transport.wa.gov.au/licensing/plates/standard/state
A2	WA DTMI, Optional Plates/Remake Application, pp.1-3; updated 30 March 2026. Dimensions, character counts and formats. Read in full.	https://transport.wa.gov.au/getmedia/d226c7d9-af3d-401d-bd0f-dd7b09d02ed2/LBU_F_Plates_Optional.pdf
A3	Queensland Government, Step up to the plate for a piece of QLD motoring history, 21 July 2020. Series-format announcement, not an issue-status forecast.	https://statements.qld.gov.au/statements/90234
N1	NIST Dataplot, Proportion Confidence Limits. Wilson/binomial interval methods; displayed update 12/11/2023.	https://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/propconf.htm
V1	Plate Recognizer, Best ANPR Camera System Setup. Engine-supplier capture guidance; live, no engine build specified.	https://platerecognizer.com/camera-setup-for-best-anpr/
V2	Plate Recognizer, LPR Camera Resolution Guide. Pixels, field of view and exposure planning; live, no engine build specified.	https://platerecognizer.com/lpr-camera-resolution-guide/
V3	Plate Recognizer, ALPR Results in the Real World. Selected-image demonstrations, including resized/rotated and altered images; not a representative site sample.	https://platerecognizer.com/alpr-results/
V4	Plate Recognizer, Stream FAQ. Effective frame rate, delay and host-processing health; live, no deployed version asserted.	https://guides.platerecognizer.com/docs/stream/help/faq/
X1	XVision, Casey Complex: Stay Time and Car Park Use. Public field-result source for the stated 2023 deployment and period.	https://xvision-ai.com/blog/case-study-casey-complex-parking-stay-time

NIST supports interval calculation, not the traffic sample design. The sample, matching and acceptance procedures are the framework defined in this paper. The illustrative calculations use their stated inputs; they are separate from Casey Complex.