Traffic Image Segmentation using U-Net

Overview

This project implements a U-Net convolutional neural network from scratch using TensorFlow for semantic segmentation of traffic scenes. The model performs pixel-wise classification to identify important road elements such as lanes and vehicles.

Problem Statement

Given a traffic scene image, the objective is to generate a segmentation mask highlighting key regions such as roads, vehicles, and lane boundaries.

Architecture

• Encoder–Decoder architecture (U-Net)
• Convolution + MaxPooling for feature extraction
• Skip connections for spatial feature preservation
• Pixel-wise binary / multi-class segmentation

Tech Stack

• Python
• TensorFlow / Keras
• NumPy
• OpenCV / PIL
• Matplotlib

Training Details

• Dataset: CamVid dataset (Kaggle)
• Images: ~700+ labeled traffic scene images
• Input Resolution: 96x128
• Epochs: 100
• Optimizer: Adam
• Loss Function: Sparse Categorical Crossentropy

Results

The model successfully segments road lanes and vehicles while preserving spatial details through skip connections.

Performance

• Training Accuracy: 94%
• Test Accuracy: 82%
• Dataset: CamVid
• Task: Pixel-wise semantic segmentation

Key Learnings

• Implemented U-Net architecture from scratch
• Built end-to-end training pipeline
• Applied data preprocessing and augmentation
• Optimized segmentation performance through hyperparameter tuning

Setup Instructions

git clone <repo-link>
cd traffic-image-segmentation
pip install -r requirements.txt
python Traffic-Image-Segmentation.py