All work
CSS 487 · Built with Manish · 2026
Road sign detection
Detects US road signs in still images and dashcam video using colour masks and shape rules, in C++ with OpenCV 4 and nothing else.
- C++
- OpenCV 4
- CMake
- sign families detected
- 5
- on dashcam video
- Live
The problem
Classic computer vision, without a neural network: find construction, guide, service, warning and regulatory signs from colour and geometry alone.
My role
Co-author with Manish.
What I built
- Colour masks for orange, green, blue and yellow sign families.
- Shape rules for diamonds, rectangles and regulatory signs such as stop and speed limit.
- The same pipeline running live on dashcam video.
The hard part
Colour thresholds that work in one frame fail in the next as light changes, so shape rules have to reject what the masks let through.
Why it matters
It shows how far careful classical vision gets before a learned model is needed, and it runs anywhere OpenCV does.
How it was measured
Run against a still-image set for each sign family and a dashcam clip.
Architecture
- 1FrameStill image or dashcam video
- 2Colour masksPer sign family
- 3Shape rulesDiamonds, rectangles, octagons
- 4DetectionsDrawn onto the frame