ParkVision 04 / 25
License-plate detection & rectification
Neural plate detector plus geometric rectification — the first two stages of the ParkVision ANPR pipeline.
Problem
Parking and access-control cameras see plates at angles, in bad light and with skew; OCR fails on raw crops.
What it does
- Finds the plate region and the vehicle box on a camera frame
- Filters candidates by confidence
- Straightens the plate (perspective and skew compensation) before OCR
How it works
- Frame from IP/parking camera or uploaded JPEG/PNG (EXIF orientation respected)
- CNN detector → candidate plate boxes → confidence filter
- Crop → geometric rectification → handed to the OCR stage
Stack & models
PythonPyTorchOpenCVONNX RuntimeCUDA
Numbers
- Whole pipeline: ~40–60 ms per frame on GPU (CUDA); CPU mode also supported
Figures come from project docs/README; no other benchmarks are claimed.