Driver Drowsiness Detection

Overview

This repository contains code and models for a machine learning and deep learning project focused on developing a drowsiness detection system. The system utilizes various image classification models to predict whether a person's eyes are open or closed, indicating alertness or drowsiness. The repository is meant to be understood in the context of the accompanying report.

Key Features

• Multiple Model Architectures: Explores the use of Convolutional Neural Networks (CNNs), Multi-Layer Perceptrons (MLPs), and Support Vector Machines (SVMs) for drowsiness detection.
• Transfer Learning: Leverages pre-trained CNN models (MobileNetV2) and fine-tuning for improved performance.
• Synthetic Data: Incorporates synthetic eye images to augment the training process and enhance model generalization.

Repository Structure

.
├── data
│   ├── CEW_Data_Test_Train   (Original, unprocessed eye image datasets)
│   └── Unity_Data_Test_Train (Cleaned, preprocessed images ready for training)
├── models
│   ├── cnn (Convolutional Neural Networks - MobileNetV2)
│   │   ├── 0_real_eyes (Models trained on real eye data)
│   │   │   ├── last_layers (Fine-tuning only the last few layers)
│   │   │   ├── last_layers (Fine-tuning only the last few layers)
│   │   │   │   ├── ckpt (Checkpoints generated during model training through keras callbacks.)
│   │   │   │   ├── history (.csv files which contain the model training history)
│   │   │   │   ├── weights (Final Model files)
│   │   │   │   └── *.ipynb (The respective Jupyter notebook)
│   │   │   └── full_network (Fine-tuning the entire network)
│   │   │   │   ├── ... (Structure similar to folder above)
│   │   ├── 1_synthetic_eyes (Models trained on synthetic eye data)
│   │   │   ├── last_layers
│   │   │   └── full_network
│   │   └── 2_fine_tuned_real_eyes (Models fine-tuned from synthetic to real)
│   │       ├── last_layers
│   │       └── full_network
│   ├── mlp (Multi-Layer Perceptron models)
│   │   ├── ... (Structure similar to CNN models)
│   └── svm (Support Vector Machine models)
│   │       ├── 0_real_eyes
│   │       └── 1_data_split
├── src
│   ├── utils.py (Data loading, preprocessing, etc.)
│   ├── plot_images.py (Several image plots)
│   └── model_evaluation.py (Evaluation scripts)
├── .gitignore
├── README.md  (This file)
└── requirements.txt

Notebooks:

• EDA&PP_*.ipynb: Performs exploratory data analysis (EDA) and pre-processing of synthetic and Unity Eyes data.
• *_real_eyes.ipynb, *_synthetic_eyes.ipynb, *_finetuned_real_eyes.ipynb: Jupyter notebooks containing the training code for each model variation.

How to Use

1. Setup:

   • Create an environment using requirements.txt: python -m venv .venv
   • Activate the environment: source .venv/bin/activate
   • Install dependencies: pip install -r requirements.txt

2. Data Preparation:

   • If the respective datasets CEW_Data_Test_Train and Unity_Data_Test_Train are available, no further action is needed.
   • Alternatively, please contact the authors for data access.

3. Model Training:

   • Open the relevant Jupyter notebooks in the models directory and adjust the repo_path variable.
   • Run the notebooks to train the different models (CNN, MLP, SVM).
   • Alternatively, use model = tf.keras.models.load_model(model_file_path, compile=True) to load any saved model from the repository in the weights folders.

4. Model Evaluation:

   • Use the model_evaluation.py script to evaluate the performance of trained models on your test dataset.