HMDDD: Hybrid Model for Distracted Driver Detection through Fusion of CNN and VGG-16

There has been an increase in the total number of road accidents leading to fatal injuries and deaths in recent years. A major reason behind these staggering figures are distracted drivers. The research aims to develop a model which will identify whether a driver is distracted or not. This study uses Convolutional Neural Networks (CNNs) to classify and detect driver distractions. Utilizing a dataset from the 2016 State Farm Challenge on Kaggle, CNN architectures were refined, including standard and modified VGG-16 models. Initial results with a simple CNN yielded a log loss of 2.67118 and an accuracy of 0.994872. Transfer learning with VGG-16 architectures improved performance, with Model Architecture-1 achieving a log loss of 1.93189. Further enhancement techniques, such as dropout, Xavier initialization, and fine-tuning, resulted in an optimized VGG-16 model with a final log loss of 1.26397, demonstrating superior accuracy.