🚗 DrowsinessNet: Real-time Drowsiness Detection

DrowsinessNet is a deep learning-based project that detects driver drowsiness from video feeds (webcam or camera). It utilizes a robust hybrid architecture combining Convolutional Neural Networks (CNN) for spatial feature extraction, Long Short-Term Memory (LSTM) networks for temporal pattern learning, and a Soft Attention mechanism to focus on the most critical frames.

🌟 Key Features

• Hybrid Architecture: CNN + LSTM + Attention Mechanism for high-accuracy temporal modeling.
• Accurate Predictions: Achieves ~93.6% accuracy and ~93.8% F1-Score on the test set.
• Real-time Processing Capability: Uses sliding window approaches with specific sequence lengths for smooth and continuous predictions.
• Web Demo (FastAPI): Includes a complete FastAPI-based web inference backend (web_app/app.py), allowing both pre-loaded demo videos and user uploads.

---

🏗️ Model Architecture

The model processes sequences of image frames (90 frames equivalent to ~3 seconds at 30 fps) to classify the state as either Alert (0) or Drowsy (1).

1. CNN Layer (Spatial Features):
   • Grayscale images (1 channel, 64x64 size) are passed through 2 Convolutional Blocks.
   • Each block uses Conv2d, BatchNorm2d, ReLU, MaxPool2d, and Dropout2d.
2. LSTM Layer (Temporal Patterns):
   • The extracted spatial features are flattened and passed to a 2-layer LSTM with 128 hidden units.
3. Soft Attention Layer:
   • A linear attention layer computes weights for each frame in the sequence, producing a weighted context vector.
4. Classification:
   • A Fully Connected (FC) layer outputs the final logic prediction (Alert vs. Drowsy).

---

⚙️ Project Structure

DrowsinessNet/
├── notebooks/                 # Notebooks for data preparation and training
│   ├── DATA PREPROCESSING .ipynb
│   └── Train.ipynb
├── models/                    # Saved trained PyTorch model weights
│   └── best_model.pth
├── web_app/                   # Web Application Backend (FastAPI)
│   ├── app.py                 # FastAPI server and routing
│   ├── model.py               # Singleton model loading & inference functions
│   ├── static/                # Static assets (CSS, JS)
│   ├── templates/             # HTML templates (index.html)
│   └── uploads/               # Temporary directory for user-uploaded videos
├── assets/                    # Sample videos and evaluation plots
│   ├── test_drowsy.mp4
│   ├── test_nondrowsy.mp4
│   ├── confusion_matrix.png
│   ├── model_flow_diagram.png
│   └── training_history.png
├── requirements.txt           # Python dependencies
└── README.md

---

📊 Training Details & Metrics

• Input size: 64x64 Grayscale
• Sequence Length: 90 frames
• Optimizer: Adam (LR: 5e-5, Weight Decay: 1e-4) with ReduceLROnPlateau scheduler.
• Loss Function: BCEWithLogitsLoss
• Evaluation Metrics (on Test Set):
  • Accuracy: 93.62%
  • Precision: 92.00%
  • Recall: 95.83%
  • F1-Score: 93.88%

---

🚀 Getting Started

1. Installation

Install the required dependencies from the requirements.txt file using pip:

pip install -r requirements.txt

2. Run the Web Demo

Navigate into the web_app directory and run the FastAPI application via Uvicorn:

cd web_app
python app.py

Or run directly with uvicorn:

uvicorn app:app --host 0.0.0.0 --port 8000

Open your browser and navigate to http://localhost:8000 to access the interactive user interface. You can upload your own mp4 videos or test with the available demo recordings.

3. Training the Model

If you wish to retrain or modify the model:

1. Put your preprocessed data appropriately inside the ../Data/train_data folder or as modified in the notebook.
2. Run all the cells in the Train.ipynb notebook to initiate training, evaluate results, and save the updated best_model.pth.