Road Pothole Detection using Supervised and Unsupervised Methods

📌 Overview

This repository contains the implementation of our Road Pothole Detection framework using Scale Invariant Feature Transform (SIFT), Bag of Visual Words (BoVW), Histogram of Oriented Gradients (HOG), and Local Binary Patterns (LBP). The project explores machine learning models like KNN, SVM, Random Forest, and Feedforward Neural Networks (FFNN) for pothole classification.

📷 System Overview

Figure 1: Systematic flowchart for road pothole detection

📜 Methodology

🔹 Feature Extraction Techniques

1. SIFT (Scale Invariant Feature Transform) + Bag of Visual Words

   • Extracts keypoints and descriptors from road images.
   • Clusters features using K-Means and transforms them with TF-IDF weighting.

2. HOG + LBP (Texture-Based Features)

   • HOG captures gradient-based features, while LBP detects texture patterns.
   • Features can be used individually or concatenated.

Figure 2: Difference of Gaussian for SIFT keypoint detection

Figure 3: Keypoint detection using 26-neighbor pixel analysis

🔹 Machine Learning Models Used

• K-Nearest Neighbors (KNN)
• Support Vector Machines (SVM)
• Random Forest Classifier (RF)
• Feedforward Neural Networks (FFNN)

📊 Results & Comparisons

🔹 Dataset

• The dataset contains 350+ pothole and plain road images.
• Images are resized to 256×256 grayscale format for uniform processing.

Figure 4: Sample images from the dataset

🔹 Classification Performance

Algorithm	Feature Used	Best Accuracy
KNN	SIFT + BoVW	77.62%
SVM	SIFT + BoVW (TF-IDF)	83.21%
SVM	HOG + LBP	92.85%
RF	HOG + LBP (Denoised)	94.28%

Figure 5: Model performance comparison

🛠 Installation & Setup

🔹 Prerequisites

• Python 3.x
• OpenCV, Scikit-learn, NumPy, Matplotlib
• TensorFlow (for FFNN model)