Traffic Sign Detection and Recognition with MATLAB

Report

See TSDR_En.pdf for a complete overview of the project.

Overview

This repository contains MATLAB code for a traffic sign and shape recognition system. The project is organized into three main pipelines:

1. Inference Pipeline (main_pipeline.m):
   Loads a test image and pre-trained models (SVM and Random Forest), performs image segmentation, extracts regions of interest (ROIs), classifies shapes using SVM, and then classifies traffic signs using a Random Forest classifier. Results and visualizations are displayed.

2. Random Forest Training Pipeline:
   Located in the RF_Model/ folder, this pipeline loads the training data from CSV files, performs feature extraction and dataset balancing, trains a Random Forest classifier (with early stopping), evaluates its performance, and saves the trained model (modelloRF.mat) in the parent folder of the current directory. The main training script is main_train.m.

3. SVM Training Pipeline:
   Located in the Shape_Classifier_SVM/ folder, this pipeline processes a large dataset of images by performing preprocessing, feature extraction, and splitting into training/test sets. It uses grid search with k-fold cross-validation to tune SVM hyperparameters, trains the SVM model, evaluates its performance, and saves the trained model (modelloSVM.mat). The main training script is main_train.m.

Repository Structure

├── main_pipeline.m           % Inference pipeline: processes a test image and performs classification using pre-trained models.
├── Shape_Classifier_SVM/     % Folder for SVM training pipeline.
│   └── main_train.m          % Main script for training the SVM classifier.
├── RF_Model/                 % Folder for Random Forest training pipeline.
│   └── main_train.m          % Main script for training the Random Forest classifier.
├── TestImages/               % Folder for test images (e.g., test1.png).
├── data/                    
│   ├── raw/                 % Raw data: CSV files (Train.csv, Test.csv, Meta.csv) and meta images.
│   └── processed/           % Processed data and results.
└── README.md                % This file.

1. Inference Pipeline (main_pipeline.m)

• Purpose:
  Processes a test image by performing color segmentation (RGB to HSI conversion), mask creation/cleaning, ROI extraction, shape classification (via SVM), and traffic sign classification (via Random Forest).

• Usage:
  Open MATLAB, navigate to the repository root, and run:

  main_pipeline.m
  

• Output:

  • Displays the original image, segmentation results, detected ROIs, and classification outcomes (including bounding boxes, labels, and scores).
  • Generates visualizations such as a montage of ROIs and confusion charts.
  • Prints a final report with classification details to the command window.

---

2. Random Forest Training Pipeline (RF_Model/main_train.m)

• Purpose:
  Loads the training dataset, extracts features, balances the dataset, trains a Random Forest classifier (with early stopping), evaluates its performance, and saves the model as modelloRF.mat (in the parent directory).

• Usage:
  Navigate to the RF_Model/ folder in MATLAB and run:

  main_train.m
  

• Output:

  • Displays training progress and evaluation metrics (e.g., training time, optimal number of trees).
  • Saves the trained model in the parent folder under the name modelloRF.mat.

---

3. SVM Training Pipeline (Shape_Classifier_SVM/main_train.m)

• Purpose:
  Processes a large dataset of images by performing preprocessing, feature extraction, and splitting into training/test sets. Uses grid search with k-fold cross-validation to determine the best SVM hyperparameters, trains the SVM model, evaluates its performance, and saves the model as modelloSVM.mat.

• Usage:
  Navigate to the Shape_Classifier_SVM/ folder in MATLAB and run:

  main_train.m
  

• Output:

  • Displays the performance of various parameter combinations and the best-found parameters.
  • Prints evaluation metrics (e.g., accuracy, F1-score) and generates confusion charts.
  • Saves the trained model as modelloSVM.mat in the current directory.

---

Output and Results

• Inference Pipeline:
  Provides visual feedback including the original image, processed masks, detected ROIs, and final classification with bounding boxes and labels.

• Training Pipelines:
  Display training progress and evaluation metrics. The trained models are saved for future inference.

---

Data Disclaimer

• Test Images: Not included. Please provide your own images in the TestImages/ folder.
• Pre-trained Models: Not included. If you do not have pre-trained models, train them using the respective training pipelines.
• Dataset Files: The dataset used is the GTSRB (German Traffic Sign Recognition Benchmark). Download the dataset from Kaggle, extract it, and copy the necessary CSV files and meta images into the data/raw/ folder.

---

Contact

For questions or further support, please contact:
Riccardo Roscica
riccardoroscica@gmail.com