الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 133 |  |  |
| | | from | 133 | | |
Abstract
Road accidents are the eighth leading cause of death for people of all ages while it is the first cause of death for children and young people. Every 24 seconds there is a person in the world who dies due to a road accident. The risk factor of road accidents increases if
the driver is distracted. One way of decreasing road accidents is to detect the different driver distractions and use this detection as an input to a driver distraction mitigation system. Here comes the role of machine learning as one of the most recent approaches
in driver distraction detection problem.
The reliability of the machine learning models depending mainly on how much the data represents the real world. Using normal RGB cameras in most of driving distractions datasets makes these datasets away from the real driving experience as RGB cameras cannot capture low lighting samples. The rest of the driver distractions datasets are
either recorded in a simulation environment or captured using IR cameras at low lighting conditions which also makes these datasets away from the real driving experience.
In order of that, the first main contribution of the thesis is presenting the largest ten-
classes driver distractions dataset to date including temporal information and low lighting support. The dataset was captured using only one sensor, a raspberry pi NoIR (No
IR filter) camera. The NoIR technology enables the camera sensor to captures RGB and
IR data at the same time combined which enables capturing data at different lighting conditions using IR LEDs. A number of 70 drivers of both genders and five different car models were involved in this dataset.
The second main contribution of the thesis is building seven End-To-End deep learning models, including three sequence models, and training them using the collected dataset. After that, we compared the seven models and analyzed the performance of each model using different techniques like confusion matrix, t-SNE representation, and saliency map. Results have shown that, among the seven models, the model with Convolutional-GRU backbone offered the best test-accuracy (95.48%). While among the non-sequence models, the model with VGG16 backbone offered the best test-accuracy (93.05%).