Back to Projects

Traffic Sign Recognition (Python Project)

Built with:

Python
TensorFlow
OpenCV
NumPy
Matplotlib

View Code

Overview

This project uses a lightweight neural network to recognize German traffic signs from low-res images. It follows a straightforward machine learning (ML) workflow: prepare, train, evaluate, and improve.

43 sign types, with extra variations added so less common signs weren't overlooked.
Repeatable training steps with results saved to outputs/.
Confusion matrix, per-class accuracy, and misclassified examples included.

Dataset

I used the German Traffic Sign Recognition Benchmark (GTSRB), a public set of photos covering 43 kinds of signs in real driving conditions. Some images are a bit blurry or partly cut off, which makes the task feel closer to real life.

Images were scaled down to 32x32 to keep the training light and fast. I also made small adjustments (tiny rotations and brightness changes) to help the model handle different lighting and angles without changing what the signs look like.

Thousands of labeled photos across 43 sign types.
Resized to 32x32; gentle edits for lighting/angle changes.
Split into training, validation, and test sets for a fair check on new images.

Source: GTSRB (official). Mirror: Kaggle.

Model Architecture

A compact CNN optimized for 32x32 inputs:

Pattern-finding layers: spot edges and shapes, then combine them into sign features; dropout step helps it learn patterns instead of memorizing photos.
Summary → decision: the network condenses each image to a short summary and then picks the best match from 43 labels.
Right-sized and stable: small enough to train quickly on 32×32 images; uses early stop/learning-rate slowdown and class weighting so rarer signs aren't overlooked.

Why this design? Simple, fast, and accurate for tiny images.

Training Setup

Short training runs, simple checks, and a few safeguards to keep learning steady.

Rounds: about 20–30 training rounds ("epochs"), stop early if it stops improving.
Batch size: 64 images at a time.
Progress check: keep a small set of images the model never sees during training to track real progress.
Learning step: slow the learning if progress stalls to fine-tune near the end.
Rare signs: give them a bit more weight so they aren’t ignored.
Image tweaks: small rotations and brightness changes during training to handle different conditions.
Saved results: charts/logs go to outputs/ in the code repo.

Results

Average (per class): 92.9% across 43 signs.
Strongest (≈100%): End of no passing by trucks; Dangerous curve (left); End of all limits; Stop; Yield.
Lower: Pedestrians (50%); End of no passing (60%); Speed limit 20 (80%).

Confusion matrix showing correct and incorrect predictions across 43 sign types — Confusion matrix — darker squares mean more correct guesses for that sign.