Driver Drowsiness & Inattention Detection

This repo contains two training notebooks (a baseline CNN and an “industry-grade” EfficientNet pipeline) and a real-time webcam inference script to classify driver behavior into 6 classes:

• DangerousDriving
• Distracted
• Drinking
• SafeDriving
• SleepyDriving
• Yawn

Project structure

• driver_drowsing_detection_model_notebook.ipynb — Baseline CNN trained from scratch; exports driving_behavior_model.h5.
• driver_drowsiness_pipeline.ipynb — Transfer-learning pipeline (EfficientNetB0 + fine-tuning, augmentation, mixed precision, richer evaluation); exports EfficientNet models.
• real_time_detection.py — Webcam inference + on-screen labels + CSV logging.
• driving_behavior_model.h5 — Trained Keras model (used by real_time_detection.py).
• driving_behavior_model.onnx — Exported ONNX version of the model (not used by real_time_detection.py as-is).
• requirements.txt — Dependency placeholder (see notes below).
• checkpoints/ — Training checkpoints (used by the pipeline notebook).

Requirements

• Windows (tested paths in notebooks are Windows-style)
• Python 3.9+ recommended
• A working webcam (for real-time inference)

Python dependencies

This project uses:

• numpy, pandas
• opencv-python
• matplotlib, seaborn
• scikit-learn
• tensorflow

Note: requirements.txt currently includes some entries that are not pip-installable packages (e.g., random, warnings, cv2, matplotlib.pyplot). If pip install -r requirements.txt fails, install the packages above directly.

Setup

1. Create and activate a virtual environment.
2. Install dependencies.

If you want a quick, reliable install, use (PowerShell):

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install numpy pandas matplotlib seaborn opencv-python scikit-learn tensorflow

Dataset

The notebooks expect the Driver Inattention Detection Dataset (as referenced in the notebook paths). You can:

• Download it to any location, then
• Update the dataset paths in the notebooks.

In driver_drowsiness_pipeline.ipynb, edit the single configuration cell:

• BASE_DIR
• TRAIN_DIR, TEST_DIR, VALID_DIR
• TRAIN_ANNOT, TEST_ANNOT, VALID_ANNOT

In driver_drowsing_detection_model_notebook.ipynb, update:

• train_images_path
• annotation_file_path
• train_folder, test_folder, valid_folder

Training

Option A — Baseline CNN (simpler)

Open and run driver_drowsing_detection_model_notebook.ipynb.

At the end, it saves:

• driving_behavior_model.h5

This file matches the default model name expected by real_time_detection.py.

Option B — EfficientNet pipeline (recommended)

Open and run driver_drowsiness_pipeline.ipynb.

It trains in two phases (frozen backbone → fine-tuning), evaluates with classification metrics, and exports:

• driver_drowsiness_efficientnet.keras
• driver_drowsiness_efficientnet.h5
• saved_model/

Real-time webcam inference

The script loads a Keras model and runs inference on frames from your webcam.

Run

From the repo folder:

python real_time_detection.py

Common options:

python real_time_detection.py --model driving_behavior_model.h5
python real_time_detection.py --webcam 0
python real_time_detection.py --save
python real_time_detection.py --normalize

• Press q to quit.
• Predictions are appended to drowsiness_predictions_log.csv.

Included model files

This repo includes two model artifacts:

• driving_behavior_model.h5 (Keras / TensorFlow) — works with real_time_detection.py.
• driving_behavior_model.onnx (ONNX) — useful for deployment in other runtimes, but the current script does not load ONNX.

If you want to run the ONNX model, you’ll need a separate inference path (typically via onnxruntime) or adapt real_time_detection.py accordingly.

Preprocessing compatibility (important)

real_time_detection.py currently preprocesses frames as:

1. BGR → RGB
2. Resize to 224x224
3. Optional /255.0 normalization via --normalize

This matches the baseline CNN notebook (which uses raw pixel values unless you explicitly normalized during training).

If you want to run the EfficientNet model exported from driver_drowsiness_pipeline.ipynb, you must ensure the webcam preprocessing matches EfficientNet’s expected preprocessing (the notebook uses tensorflow.keras.applications.efficientnet.preprocess_input, which scales to roughly [-1, 1]).

Practical options:

• Use the baseline driving_behavior_model.h5 for real_time_detection.py as-is, or
• Update real_time_detection.py to apply EfficientNet preprocessing before prediction.

Outputs

• On-screen overlay: predicted class, confidence, FPS
• CSV log: drowsiness_predictions_log.csv with columns timestamp, iso, class, confidence
• Optional video output: output_with_labels.avi (when using --save)

Troubleshooting

• Webcam won’t open: try a different --webcam index (0, 1, 2…). Close other apps using the camera.
• Model file not found: pass an explicit path via --model.
• TensorFlow install issues on Windows: ensure you’re using a supported Python version and upgraded pip.
• Wrong predictions / low confidence: verify that preprocessing at inference matches the preprocessing used during training (normalization and/or EfficientNet preprocess_input).

Notes

• The checkpoints/ directory is used by the EfficientNet pipeline notebook to save best models during training.
• Class ordering must stay consistent between training and inference. real_time_detection.py uses: ['DangerousDriving','Distracted','Drinking','SafeDriving','SleepyDriving','Yawn'].

License

Add a license if you plan to distribute this project.