Saransh-cpp/IIITD_ALIVE_DSM

Submission for IIITD's ALIVE project round 2.

IIITD ALIVE DSM submission

Submission for IIITD's ALIVE project round 2.

Overview

Relevant files

• train.ipynb
• drowsiness_detector.py
• model.tflite

Dataset

The model uses a combination of 2 datasets to improve the accuracy. These datasets are -

• https://www.kaggle.com/prasadvpatil/mrl-dataset (2000 open eye images and 2000 closed eye images)
• https://www.kaggle.com/serenaraju/yawn-eye-dataset-new (Dataset used partially, i.e, only the closed and open eye images were used. 726 open eye images and 726 closed eye images)

The total number of images used -

• 2726 open eye images
• 2726 closed eye images

Usage

1. Clone this repository.

git clone https://github.com/Saransh-cpp/IIITD_ALIVE_DSM

2. Create a virtual environment

cd IIITD_ALIVE_DSM
python -m venv .env

3. Activate the environment

.env/Scripts/activate

4. Install the requirements

pip install -r requirements.txt

5. Run train.ipynb to re-train the model (the model is already trained).
6. To start the live video feed for drowsiness detection, run -

python drowsiness_detector.py

Note: You might need to edit thhe number passed in VideoCapture to your webcam's number (0, 1, 2, 3, ....) -

IIITD_ALIVE_DSM/drowsiness_detector.py

Line 26 in ced24be

cap = cv2.VideoCapture(3)

7. To stop the live feed press q.

Model architecture

• The training step uses Transfer Learning with VGG19.
• Last layer of VGG19 has been removed and a Flatten layer with a new output layer has been added.
• The model uses adam optimizer and categorical_crossentropy loss function.

The train notebook is very well documented.

Drowsiness setector in action (screenrecording)

2022-01-11.23-17-22.mp4

Extras

The repository contains a CI built using GitHub Action which can be scaled up to include testing of the scripts.

Tasks

Task 1

The model has been trained and converted into tflite format in the train notebook. The trained tflite model is also present here.

Task 2

Drowsiness detector to run on live video feed has been created in drowsiness_detector.py.