Sandeep2803/traffic-segmentation-alexnet

🚗 Traffic Scene Semantic Segmentation using a Custom AlexNet-based Encoder–Decoder Architecture

A PyTorch implementation of semantic segmentation using a custom AlexNet-based encoder–decoder architecture trained on the BDD100K dataset.

📌 Project Overview

This project implements a Semantic Segmentation model using a custom AlexNet-based Encoder–Decoder Architecture in PyTorch. The model is trained on the BDD100K dataset to perform pixel-wise classification of traffic scene images.

Unlike image classification, semantic segmentation predicts a class label for every pixel in an image. The model identifies different objects such as roads, vehicles, buildings, vegetation, sky, pedestrians, traffic signs, and other traffic scene elements.

The objective of this project is to understand and implement the complete semantic segmentation pipeline, including dataset preprocessing, model development, training, evaluation, visualization, and performance analysis.

✨ Features

• —Custom AlexNet Encoder
• —Custom Decoder for Semantic Segmentation
• —Pixel-wise Image Segmentation
• —BDD100K Dataset Support
• —Custom Dataset and DataLoader
• —Training Pipeline
• —Evaluation Pipeline
• —Visualization of Predictions
• —Color Segmentation Masks
• —Pixel Accuracy Evaluation
• —Mean Intersection over Union (mIoU)
• —Training Loss Visualization

🗂 Dataset

Dataset Name: BDD100K (Berkeley DeepDrive)

BDD100K is a large-scale autonomous driving dataset containing real-world road scenes with pixel-level semantic segmentation annotations.

Classes

The model predicts 19 semantic classes, including:

• —Road
• —Sidewalk
• —Building
• —Wall
• —Fence
• —Pole
• —Traffic Light
• —Traffic Sign
• —Vegetation
• —Terrain
• —Sky
• —Person
• —Rider
• —Car
• —Truck
• —Bus
• —Train
• —Motorcycle
• —Bicycle

🏗 Model Architecture

The semantic segmentation model uses a custom AlexNet-based Encoder–Decoder architecture.

[image]

📂 Project Structure

Traffic-Segmentation-AlexNet/

│
├── models/
│   ├── encoder.py
│   ├── decoder.py
│   └── segmentation_model.py
│
├── preprocessing/
│   ├── dataset.py
│   ├── dataloader.py
│   └── transforms.py
│
├── outputs/
│   ├── predictions/
│   ├── plots/
│   └── metrics.txt
│
├── config.py
├── train.py
├── test.py
├── evaluate.py
├── metrics.py
├── visualize.py
├── plot_loss.py
├── utils.py
├── requirements.txt
└── README.md

⚙ Training Configuration

📊 Final Results

🖼 Sample Prediction

Example 1

[image]

Example 2

[image]

Example 3

[image]

Example 4

[image]

Example 5

[image]

🚀 Installation

Clone the repository

git clone https://github.com/Sandeepkumarreddy-7/Traffic-Segmentation-AlexNet.git

Move into the project

cd Traffic-Segmentation-AlexNet

Create Virtual Environment

python -m venv venv

Activate Virtual Environment

Linux

source venv/bin/activate

Windows

venv\Scripts\activate

Install Dependencies

pip install -r requirements.txt

▶ Training

python train.py

📈 Evaluation

python evaluate.py

🖼 Testing

python test.py

📉 Plot Training Loss

The following graph shows the decrease in training loss during model training.

[image] ---

🛠 Technologies Used

• —Python
• —PyTorch
• —TorchVision
• —NumPy
• —Matplotlib
• —Pillow
• —Git
• —GitHub

📈 Future Improvements

• —Use Pretrained AlexNet Weights
• —Add Skip Connections
• —Implement Data Augmentation
• —Train for More Epochs
• —Use Dice Loss or Focal Loss
• —Improve Decoder Architecture
• —Deploy with Gradio or Streamlit
• —Experiment with U-Net and DeepLabV3+

👨‍💻 Author

Dammuru Sandeep Kumar Reddy

B.Tech – Computer Science and Engineering

Srinivasa Ramanujan Institute of Technology

GitHub: https://github.com/Sandeepkumarreddy-7

⭐ Acknowledgements

• —Berkeley DeepDrive (BDD100K)
• —PyTorch
• —TorchVision
• —OpenAI ChatGPT