Arseni Aliakseichyk
Embedded and Infrastructure Engineer
Real-Time Shape & Symbol Classifier
Edge ML Pipeline: PyTorch → ONNX → Raspberry Pi 5
Academic • 2025A complete machine learning pipeline built for a university project - from dataset creation and model training to real-time edge inference. I trained a MobileNetV3 model in PyTorch on ~16,000 synthetic images across 80 classes (geometric shapes, mathematical symbols, arrows, etc.), exported it to ONNX format, and deployed it on a Raspberry Pi 5 where it runs inference at 25–30 FPS on 640×480 camera input. The system includes a custom SPI LCD driver for the ST7735S display (RGB565 pixel format via libgpiod) and an HTTP dashboard for real-time classification statistics and confusion matrix visualization.
Key Features
- MobileNetV3 trained on ~16K synthetic images (80 classes)
- ONNX export and optimized edge inference at 25–30 FPS on RPi 5
- OpenCV camera capture with dynamic ROI and configurable confidence threshold
- Custom SPI LCD driver (ST7735S, RGB565 via libgpiod)
- HTTP stats API with live confusion matrix and per-class accuracy
- Dataset augmentation pipeline with rotation, noise, and scaling