US Traffic Sign Classification

This project is a complete machine learning pipeline for classifying US Traffic Signs from the LISA dataset using deep neural network embeddings and traditional/custom classifiers.

Dataset:

LISAC Traffic Sign Classification Dataset

Overview

Processes raw images, extracts high-level semantic features using MobileNetV2, reduces dimensionality using PCA and MDA, and evaluates multiple classifiers and Bayesian classifiers (Parametric and Non-Parametric).

Requirements & Setup

1. Clone the repository and navigate into the project directory.
2. Create and activate a virtual environment.
3. Install the required dependencies:

pip install -r requirements.txt

How to run

python run_pipeline.py

This script will run:

1. preprocess.py: Splits the raw dataset into Train/Valid/Test.
2. extract_features.py: Extracts 1280-dimensional features using MobileNetV2.
3. PCA_dim_reduction.py & MDA_dim_reduction.py: Reduces dimensions using PCA and MDA.
4. train.py: Trains SVM, Random Forest, k-NN, and Bayesian Classifiers.
5. evaluate.py: Generates confusion matrices, reports, and error.
6. repeated_experiments.py: Evaluates variance and stability of classifiers via multiple data splits.

Summary of Results

Average test set accuracy and stability metrics achieved across 5 randomized training splits:

PCA (50 Dimensions)

Classifier	Mean Accuracy	Variance	Std Dev	Mean Time (s)
SVM	91.61%	0.000027	0.0052	0.1172
k-NN	92.34%	0.000027	0.0052	0.2569
Bayesian Non-Parametric	92.34%	0.000027	0.0052	0.0754
Bayesian Parametric	90.66%	0.000009	0.0029	0.0055
Random Forest	86.57%	0.000167	0.0129	0.5972

MDA (c-1 Dimensions)

Classifier	Mean Accuracy	Variance	Std Dev	Mean Time (s)
k-NN	91.02%	0.000030	0.0055	0.0068
Bayesian Non-Parametric	90.58%	0.000061	0.0078	0.0224
SVM	90.36%	0.000056	0.0075	0.0221
Bayesian Parametric	89.12%	0.000061	0.0078	0.0014
Random Forest	82.55%	0.000556	0.0236	0.0990

---

Sample images per classes:

---

Aashish Harishchandre