Overview
A weekend research prototype, not a product. YOLOv8 fine-tuned for aerial license plates, PaddleOCR reading plates on the fly, fuzzy matching against a payment API, and a three-thread reader → inference → display pipeline holding ~30 FPS, with a Streamlit dashboard. The hard part was not the ML, it was keeping the pipelines from fighting each other.
The problem
Parking enforcement still relies on manual patrols. What if a drone could patrol a lot, scan every vehicle and flag unpaid parking within seconds?
How it was built
YOLOv8 fine-tuned for license plates seen from an aerial perspective, PaddleOCR reading plates on the fly, and fuzzy matching against the parking system's list of paid vehicles. A three-thread pipeline, reader → inference → display, keeps the feed near 30 FPS. A Streamlit dashboard shows live telemetry, plate insights and an analysis log.
What it does
- 01Real-time vehicle and plate detection from drone footage
- 02Plate OCR with confidence scoring tuned to feel reliable rather than noisy
- 03Cross-reference against a payment API; any visible vehicle not on the paid list is flagged
- 04Dark-mode dashboard with live telemetry
What I took from it
The hardest part was not the machine learning. It was the engineering around it: keeping OCR, detection and UI pipelines from fighting each other, and making confidence scores trustworthy. Real-time systems expose weak links quickly. If you are working on something similar, email me; I would rather compare notes than publish half a prototype.