Personal Accomplishments
- ML Strategy:
- Drafted the following edge + cloud ML flow: Frame Acquisition -> Frame Preprocessing -> License Plate Detection -> Transfer Cropped License Plate to Cloud -> License Plate Preprocessing -> License Plate Classification -> OCR -> Text Formatting
- ML Dataset Selection:
- Identified a license plate detection dataset from Kaggle
- Web-scraped platesmania.com to build a US license plate OCR dataset, as an ideal dataset wasn’t readily available
- Chose the OpenALPR dataset for end-to-end testing
- ML Model Selection:
- Benchmarked FastALPR from GitHub on RPi 4
- Decided to fine-tune YOLOv11n for plate detection
- Chose to train an OCR model specifically for US license plates
- Hardware Selection:
- Based on latency benchmarks, finalized the dash cam hardware:
- Raspberry Pi 5 as the SBC
- Raspberry Pi Global Shutter Camera with a 6mm wide-angle CCTV lens for improved field of view
- Adafruit GPS breakout module for geotagging plate captures
- Based on latency benchmarks, finalized the dash cam hardware:
Progress
My progress is on schedule.
Schedule
- Draw out the technical block diagram for the dash cam
- Order dash cam components
- Augment US license plate dataset
- Prepare the design presentation and report