A readable plate image, correct OCR text and a correct gate decision are three different results. Plan and test each stage. A parking overview that shows a vehicle may lack usable plate detail; a correct plate read can still lead to a wrong access decision if a list is expired or the controller receives a duplicate event.
Separate capture, recognition and authorization
| Stage | Output | Acceptance question |
|---|---|---|
| Plate capture | Saved frame with legible characters | Can a reviewer read the complete plate from the original recording? |
| OCR / recognition | Plate string, time and relevant event metadata | Does the software return the correct full string for local plate formats? |
| Access decision | Permit, refuse or manual review | Does the correct permission apply to this lane, time and plate? |
LPR and ANPR commonly describe automatic plate-reading workflows; request the actual functions rather than choosing by acronym. Recognition may run inside a compatible camera, on a server or in another analytic service. Ordinary NVR video recording does not establish OCR, plate search or barrier integration. Ask for software licenses, compatible devices, supported countries/formats and the destination for events and saved images.
Design the reading zone before choosing the lens
Record lane width, approach direction, reading-zone distance, camera height and sideways offset, plate position, required vehicle types and maximum expected speed. Decide whether vehicles stop at a barrier or move freely. Check front/rear plates and motorcycles where relevant; a solution designed for one plate layout may not read another.
Measure plate width in pixels on the original saved frame at both edges of the reading zone. A wider scene spreads the available pixels across more road. Digital zoom enlarges existing pixels and does not recreate lost character detail. Check focus and geometry at the intended zone, not only on a vehicle parked nearer the camera.
Axis's plate-capture white paper explains the effects of scene width, illumination, motion and viewing angle. Its recommendations belong to its described equipment and conditions. For a specific third-party example, the AXIS License Plate Verifier manual specifies at least 130 pixels across a one-row plate or 70 across a two-row plate, mounting angles no larger than 30° in any direction and limited roll. These are that application's criteria, not universal limits or QuarkView specifications. Use the selected recognition product's manual for your design.
Balance speed, shutter and lighting
Motion blur depends on how far plate detail moves across the image during exposure. Shorter exposure can preserve characters but gives the sensor less light. At night, inspect actual plates with headlights on; a view that looks bright overall can have a washed-out plate. Tune the plate view for legibility and retain an overview view for vehicle context.
Check focus after dark, lighting at the near and far ends of the zone, reflective plate saturation and shadows from vehicle geometry. Test the actual plate materials and local layouts. Avoid assuming that a general night-vision range is an LPR range. Changing exposure or illumination to solve one failure should be followed by tests at other positions and speeds.
| Failure | Likely area to investigate | Repeat test |
|---|---|---|
| Too few pixels / uncertain characters | Scene width, lens and reading distance | Compare original-frame plate width before and after narrowing the view |
| Trailing or smeared strokes | Motion, exposure and insufficient light | Pass at recorded speeds using shorter exposure with appropriate illumination |
| Bright blank plate at night | Exposure and reflective plate illumination | Compare headlights and illumination levels without losing character contrast |
| Readable image, wrong string | Country/format support, OCR settings and ambiguous characters | Compare software output with known full plate text |
| Second vehicle missed | Occlusion, reading zone or event suppression | Test separated vehicles and closely following vehicles safely |
| Correct text, wrong gate action | Permission rules, event mapping and stale lists | Inspect the OCR event and controller log for the same passage |
Example: a controlled single entry lane
Planning example, not a field result: assume a 3 m-wide one-way lane, a marked reading zone 6 m from the camera, a 10 km/h operating speed limit and local one-row plates. Use a dedicated plate view, a separate overview view, recording and a verified recognition application. Document the chosen application's required plate size, supported speed and illumination before selecting hardware. The 6 m distance and 10 km/h limit are site assumptions, not supported specifications for a QuarkView model.
Pixel-budget illustration: assume a 1,920-pixel-wide image shows exactly 3.0 m at the plate plane and a front-facing plate is 0.52 m wide. Ideal projected width is 1,920 × 0.52/3.0 ≈ 333 pixels. Widening that scene to 6.0 m reduces it to about 166 pixels. This simple proportional estimate assumes a flat, front-facing target at the measured plane; perspective, angle, cropping and lens distortion change the result. Measure actual frames and compare with the selected software's requirement. Pixel count alone does not establish readable focus or OCR success.
Motion illustration: 10 km/h is about 2.78 m/s. During 1/1,000 s exposure a vehicle travels about 2.8 mm; during 1/100 s it travels about 27.8 mm. That is physical travel, not image blur in pixels: direction, angle and magnification determine projected movement. This explains why a shutter setting needs to be tested with actual speed and light rather than copied as a universal recipe.
Count errors by passage and condition
- Agree the test set and success criteria before the trial. Log every passage with ground-truth plate, lane/direction, speed, day/night condition, saved frame, OCR output and controller action.
- Count readable saved plates against all passages. Keep dirty, hidden or absent plates as labelled failures or exclusions defined in advance, not silently removed successes.
- Count exactly correct full-string reads, missing reads and wrong reads separately against the same passage total. If reporting OCR success only among readable captures, label that conditional denominator too.
- Check timestamps, direction, duplicates and association with the overview recording. Record several events for one passage as duplicates rather than several successful vehicles.
- Repeat across headlights, normal local formats, permitted vehicle speeds and weather conditions that matter to the site. Mark untested conditions explicitly.
- Review the saved image for each wrong or absent result before changing one setting and rerunning comparable passes.
Reporting example only: assume 100 test passages, 92 readable saved plates, 86 correct OCR strings, 6 wrong strings and 8 absent reads. Capture yield is 92/100 = 92%; exact-read yield is 86/100 = 86%. If all 86 correct reads are within the 92 readable images, conditional recognition yield is 86/92 ≈ 93.5%. Both can be reported, but 93.5% alone hides eight capture failures. These invented counts are not a product rating or a recommended acceptance threshold. Keep day and night subsets separate so the total cannot hide a poor night result.
Make the access decision independently testable
Define the event interface with the gate installer: plate text and normalization, timestamp/time zone, direction, lane, unique event identifier and relevant confidence or status. Document where the authorized list resides and how changes, expirations and revocations propagate. Test reconnection and duplicated messages so one vehicle does not produce unintended repeated actions.
Exercise authorized, expired, unlisted and deliberately similar plates; verify the controller's final log, not just the camera's recognition screen. Test a network loss, recognition service restart and stale-list condition. Agree with the access and gate-safety installer what happens when the system cannot decide, and provide a controlled manual procedure. OCR must not override the controller's safety logic.
A recognized number is not proof of the driver's identity or a genuine credential: plates can be copied and a closely following vehicle may share an open gate cycle. For higher-consequence access, evaluate an additional credential and anti-tailgating measures with the installer rather than treating OCR as the entire access system.
Manage movement records and list ownership
For UK organisations, ICO ANPR guidance calls for a justified purpose and camera placement, a DPIA, prominent notices, accurate matching and retention linked to need. It treats vehicle registration marks as personal data in most relevant ANPR circumstances. That UK guidance is not a rule for every jurisdiction.
Assign separate responsibility for plate permissions, correction of wrong numbers, video retrieval, export and deletion. Set access to search logs and reference lists as well as video. A short video retention period does not automatically delete a separate plate-event database. Review revocations and repeat the lane test after a moved camera, lens change, updated software or changed lighting.
Include a lane drawing, distance, speed range, local plate formats, sample conditions and existing controller/software in your LPR requirement inquiry. Use the parking overview guide for wider scene coverage and vehicle detection guide for event alerts that do not need a plate number.