DSM

The DSM system detects whether a driver of an automobile is feeling sleepy using a camera and TensorFlow-lite model.

The training notebook is Train_NB.ipynb

Model directory is image_classification

The problem statement

Given was a pre-cleaned and pre-normalized dataset from Kaggle which had just two classes - Open eyes and Closed eyes.

The way how this helps us in creating a drowsiness detector is by the fact that drowsiness may be defined by an action that involves closing eyes due to the fatigue of the driver.

Usage for inference

1) Make sure your computer is connected to a webcam. 2) Go to the directory of the model - 'tree/master/lite/examples/image_classification/raspberry_pi' (cd ../tree/master/lite/examples/image_classification/raspberry_pi) 3) Run requirements.txt to install required libraries using pip install -r requirements.txt 4) After installation, run python classify.py --model eyemodel1_metadata.tflite to get real-time video inference

Proof of concept

When driver is active

When driver is slightly drowsy

When driver starts sleeping on the wheel

DSM Model architecture

The initialized keras sequential model consisted of:

1. data_augmentation -- Data Augmentation layer to apply Random Flip, Random Rotation, and Random Zoom effect on images
2. Rescaling(1./255) -- Rescale images
3. Conv2D(16, 5, padding='same', activation='relu' -- Convolution layer with 5x5 strides and 16 filters activated with Rectified Linear Unit activation function
4. MaxPooling2D(), -- Pool maximum values of individual pixels and reduce size by half
5. Conv2D(32, 5, padding='same', activation='relu'), -- Convolution layer with 5x5 strides and 32 filters activated with Rectified Linear Unit activation function
6. MaxPooling2D(), -- Pool maximum values of individual pixels and reduce size by half
7. Conv2D(64, 3, padding='same', activation='relu'), -- Convolution layer with 3x3 strides and 64 filters activated with Rectified Linear Unit activation function
8. MaxPooling2D(), -- Pool maximum values of individual pixels and reduce size by half
9. Conv2D(64, 3, padding='same', activation='relu'), -- Convolution layer with 3x3 strides and 64 filters activated with Rectified Linear Unit activation function
10. MaxPooling2D(), -- Pool maximum values of individual pixels and reduce size by half
11. Conv2D(64, 3, padding='same', activation='relu'), -- Convolution layer with 3x3 strides and 64 filters activated with Rectified Linear Unit activation function
12. MaxPooling2D(), -- Pool maximum values of individual pixels and reduce size by half
13. Conv2D(64, 5, padding='same', activation='relu'), -- Convolution layer with 5x5 strides and 64 filters activated with Rectified Linear Unit activation function
14. MaxPooling2D(), -- Pool maximum values of individual pixels and reduce size by half
15. Dropout(0.2), -- Randomly deactivate/dropout 20% of neurons to normalize and help in reducing overfitting
16. Flatten(), -- Flatten tensors
17. Dense(640, activation='relu'), -- 640 dense layer neurons with Rectified Linear Unit activation to simplify weight distribution
18. Dense(2) -- 2 neurons because of 2 classes in the dataset. Acquires probability of both the function.

Optimizer: Adam optimizer to adjust learning rate as per the momentum required

Loss function - SparseCategoricalEntropy: Entropy across the network is acquired across different classes. Mostly used in Multiclass classification but gives good accuracy in Binary Classification problems too.

Built With

• jupyter-notebook
• python
• shell