Computer Vision Developer in Pakistan
Computer vision gets hard when it leaves the lab: glare, motion blur, dirty plates and text formats no public dataset covers. I work on that end of the problem, using YOLO for detection and PaddleOCR for recognition, then deploying the result on edge hardware.
The two-stage approach matters. Detection finds the region of interest, the crop removes background noise, and a smaller recognition model reads only what is left. Each stage can be measured and improved separately.
Tuning for local data
Generic OCR models stumble on non-standard local number plate layouts. I collected custom data, labelled it, and fine-tuned PaddleOCR on it, which is why the SmartPark pipeline reaches 90%+ OCR accuracy on formats a stock model handles poorly.
Edge constraints shape model choice
On a Raspberry Pi every millisecond and megabyte counts. I profile models early, convert ONNX to NCNN for about 3x faster inference, and use an event-driven capture strategy of exactly three frames per trigger instead of continuous video.
Relevant case studies
SmartPark
Sensor trigger to on-device plate recognition to a real barrier opening, with ~99% detection accuracy.
Read the SmartPark case study
Frequently asked questions
Which frameworks do you use?
YOLO for detection, PaddleOCR for text recognition, ONNX and NCNN for deployment, CVAT and LabelOCR for labelling.
Do you need a large dataset?
Not always. Fine-tuning and model-assisted pre-labelling reduce the amount of hand-labelling needed, but real images from your site are the most valuable input.
Can you deploy on a Raspberry Pi?
Yes. That is exactly what SmartPark does, with GPIO control of a physical barrier.
Let's talk about your project
Email dev@abdullahaltaf.com with a few lines about the product, the timeline and the role you have in mind, or use the contact form on the homepage.