PlateTracker.

// overview

PlateTracker reads license plates from live video. A custom-trained YOLOv8 model finds plates frame by frame, every plate gets a persistent track, and an OCR stage reads the characters. A plate only counts once several reads agree. A second iteration, LicenceRIP, swapped the OCR engine for FastALPR configured for European and Croatian plates and compared the two.

• Python
• OpenCV
• YOLOv8
• PaddleOCR
• FastALPR
• NumPy

// problem

OCR on a single video frame is noisy. Plates are small, blurred and seen at an angle, and one bad frame turns ZG 1234-AB into ZG 1Z34-A8. Logging every raw read would bury the real plates under near-duplicates and errors, so the system has to decide when a reading is actually trustworthy, and it has to do it fast enough to keep up with the video.

01

Detect

A YOLOv8 model trained on plate data finds plates in every frame. Detections below a minimum pixel size are skipped, since they are too small to read reliably anyway.

02

Track

Each plate gets a track ID, so readings from different frames are grouped by vehicle instead of treated independently. History is capped at 120 frames per track to bound memory.

03

Read

OCR runs every second frame per track instead of every frame. Crops are trimmed by a small margin and upscaled 3× first, so distant plates still have enough pixels for the OCR to work with.

04

Confirm

Candidate reads are voted on across frames. A plate is confirmed after two consistent reads, or straight away at 0.8 confidence or more, and reads within one character of each other are merged into the same candidate.

05

Log

Confirmed plates are written once to a timestamped log. A debug view shows every crop the OCR sees, which made tuning the thresholds against real footage much faster.

// highlights

• +Two OCR backends compared on the same pipeline: PaddleOCR (angle classification, GPU) and FastALPR tuned for EU/HR plates
• +Multi-frame voting with a one-character similarity threshold to suppress OCR noise
• +OCR throttled to every second frame per track to keep the loop real-time
• +Every threshold in one config block: crop margin, upscale factor, confidence, history length

// what I learned

• →Most of the accuracy came from what happens around the model (cropping, upscaling, voting), not from the model itself.
• →Tracking turns a per-frame problem into a per-object one, and that is what makes reasoning over time possible.
• →Making every threshold explicit config made tuning a matter of minutes instead of code changes.