🛣️ SafeDrive AI: Road Damage Detection System

A Deep Learning project designed to enhance road safety by automatically detecting and classifying road surface damages (Cracks and Potholes) using Computer Vision.

📌 Project Overview

Road maintenance is critical for vehicle safety and infrastructure longevity. This project implements a Convolutional Neural Network (CNN) to:

• Automate the inspection of road surfaces 🚗
• Classify road conditions into three categories: Clean, Crack Damage, and Pothole 🛠️
• Provide real-time-ready classification using the lightweight MobileNetV2 architecture ⚡

🚀 Features

• 🖼️ Image Classification: High-accuracy detection of road defects.
• 🧠 Transfer Learning: Leveraging MobileNetV2 pretrained on ImageNet for superior feature extraction.
• ⚡ GPU Accelerated: Optimized for training on NVIDIA T4 GPUs.
• 📊 Performance Tracking: Real-time monitoring of accuracy and loss during 20 training epochs.
• 🧪 Inference Script: A dedicated cell for testing the model on new, unseen road images.

🧰 Technologies Used

• Python 🐍
• TensorFlow & Keras (Deep Learning Framework)
• MobileNetV2 (Base Model Architecture)
• Matplotlib (Result Visualization)
• Google Colab (Cloud Computing & GPU)

📂 Dataset Structure

The dataset is organized into three main categories, split into Training, Validation, and Testing sets:

1. Clean: Well-maintained road surfaces.
2. Crack Damage: Surfaces showing longitudinal or transverse cracks.
3. Pothole: Significant road depressions requiring immediate repair.

📈 Model Performance

After training for 20 Epochs, the model achieved outstanding results:

• Training Accuracy: 99.13% ✅
• Validation Accuracy: 98.41% ✅
• Training Loss: 0.0295
• Validation Loss: 0.0428

🧠 Project Workflow

1. Data Preprocessing: Image rescaling (1/255) and target sizing (224x224).
2. Architecture Design: Global Average Pooling followed by a Dropout layer (0.3) to prevent overfitting.
3. Model Training: Using Adam optimizer and Categorical Crossentropy loss.
4. Evaluation: Testing the model on a separate k_t folder to ensure real-world reliability.

💡 Future Improvements

• Transition to YOLOv8 for real-time Object Detection (Bounding Boxes).
• Expand the dataset to include diverse weather conditions (Rainy/Night).
• Deploy the model as a mobile application for road maintenance teams.

👨‍💻 Author

Fahd

⭐ Project Goal

To provide an AI-driven solution that assists municipalities and transport authorities in identifying road hazards quickly and efficiently.