Traffic Light Classification with ResNet-50

Overview

This project implements a traffic light classification model using the ResNet-50 architecture to classify traffic light states into three classes: Go, Stop, and Warning. The model is trained on the LISA Traffic Light Dataset, which contains annotated images of traffic lights captured in various conditions.

Dataset

The LISA Traffic Light Dataset from Kaggle is used for training and evaluation. Below are the detailed characteristics of the dataset:

• Source: LISA Traffic Light Dataset by Mikkel Bornoe.
• Content:
  • Images: Approximately 43,007 images with 113,888 annotated traffic lights captured in urban environments.
  • Sequences: The dataset is divided into day and night sequences:
    • daySequence1: Daytime images with annotations for traffic light states.
    • daySequence2: Additional daytime images with similar annotations.
    • dayClip1 to dayClip13: Short video clips (extracted frames) capturing daytime traffic light sequences.
    • nightSequence1: Nighttime images with annotations.
    • nightSequence2: Additional nighttime images.
    • nightClip1 to nightClip5: Short video clips (extracted frames) capturing nighttime traffic light sequences.
  • Annotations: Provided in CSV files, including:
    • Filename of the image.
    • Bounding box coordinates (x_min, y_min, x_max, y_max) for traffic lights.
    • Class labels: go, stop, warning, or none (for images without a clear traffic light state).
  • Classes: The project focuses on three classes:
    • Go: Green light, indicating vehicles can proceed.
    • Stop: Red light, indicating vehicles must stop.
    • Warning: Yellow light, indicating caution or preparation to stop.
  • Image Format: PNG files.
  • Resolution: Varies (typically 1280x960 pixels).
  • Conditions: Images capture diverse conditions, including different lighting (day, night), weather, and angles.
• Size: The dataset occupies approximately 5.7 GB when extracted.
• Directory Structure:

  dataset/
  ├── daySequence1/
  ├── daySequence2/
  ├── dayClip1/
  ├── ...
  ├── dayClip9/
  ├── nightSequence1/
  ├── nightSequence2/
  ├── nightClip1/
  ├── ...
  ├── nightClip5/
  ├── Annotations/
  │   ├── daySequence1.csv
  │   ├── daySequence2.csv
  │   ├── nightSequence1.csv
  │   ├── nightSequence2.csv
  │   ├── dayClips.csv
  │   ├── nightClips.csv
  

• Usage: The dataset is split into training, validation, and test sets during preprocessing. Images are resized to 224x224 pixels to match ResNet-50 input requirements.

Download the dataset from Kaggle and extract it into the data/ directory.

Install dependencies using:

pip install tensorflow numpy pandas opencv-python matplotlib scikit-learn

Project Structure

traffic-light-classification/
│
├── lisa_raw/                  # Directory for extracted dataset
├── processed/                 # Directory for cropped& preprocessed dataset
├── dataset/                   # Directory for Ready to use dataset
├── models/                    # Directory for save trained models
├── visuals/                   # Directory for saved plots& curves
├── notebooks/                 # Directory for Jupyter notebooks for data exploration and analysis
├── README.md                  # Project documentation

Setup

1. Clone the Repository:

   git clone https://github.com/3omd4/traffic-light-classifier.git
   cd traffic-light-classification

2. Download the Dataset:

   • Download the LISA Traffic Light Dataset from Kaggle.
   • Extract the dataset into the data/ directory.

3. Install Dependencies:

   pip install -r requirements.txt

Usage

1. Preprocess the Data: Run the preprocessing script to prepare the dataset for training:

   python src/preprocess.py

   This script resizes images, normalizes pixel values, and splits the data into training, validation, and test sets.

2. Train the Model: Train the ResNet-50 model using:

   python src/train.py

   The script loads the preprocessed data, fine-tunes ResNet-50, and saves the trained model to the models/ directory.

3. Evaluate the Model: Evaluate the model's performance on the test set:

   python src/evaluate.py

   This generates metrics such as accuracy, precision, recall, and a confusion matrix.

Model Details

• Architecture: ResNet-50 (pre-trained on ImageNet, fine-tuned for traffic light classification).
• Input: Images resized to 224x224 pixels.
• Output: Classification into three classes: go, stop, warning.
• Training:
  • Optimizer: Adam
  • Loss: Categorical Crossentropy
  • Batch Size: 32
  • Epochs: 50
  • early_stopping: patience=5

Model Versions

• Version 1:

  • Only Using: daySequence1 and daySequence2.
  • Description: The initial model was trained on a subset of the dataset, focusing on daytime sequences. It achieved moderate performance but was limited by the smaller dataset size and lack of diversity in lighting conditions.

• Version 2:

  • Only Using: daySequence1, daySequence2, dayClip1, dayClip2, dayClip3, dayClip4, dayClip5.
  • Description: The improved model incorporates additional daytime clips, significantly increasing the dataset size and diversity. This led to a substantial improvement in accuracy, making the model more robust for daytime scenarios.

Training Curves

The following plots show the model's performance during training:

Contributing

Contributions are welcome! Please:

1. Fork the repository.
2. Create a new branch (git checkout -b feature-branch).
3. Submit a pull request with a clear description of changes.

License

This project is licensed under the MIT License. See the LICENSE file for details.

Acknowledgments

• LISA Traffic Light Dataset by Mikkel Bornoe.
• TensorFlow/Keras for providing the ResNet-50 implementation.