Detecting Distracted Drivers Using Convolutional Neural Networks

Driver distraction has become a major concern for road safety as distracted driving is a leading cause of car accidents, resulting in a large number of injuries and fatalities every year. This project aims to address the issue of driver distraction as a leading cause of traffic accidents worldwide. Utilizing various image datasets, the project explores the use of Convolutional Neural Networks (CNNs), specifically the DarkNet53 model, to classify whether drivers are focused on the road or distracted. A collection of publicly available datasets and a new dataset consisting of 4024 images produced by the authors were used in training the model, with image augmentation being used to significantly enhance the model performance. The training and testing processes were conducted using MATLAB and different techniques such as transfer learning and layer freezing were utilized to improve the training process. The real-time performance of the DarkNet53 model was also evaluated against other models and it was shown that the DarkNet53 had adequate performance while maintaining a high level of accuracy. Furthermore, the model misclassifications were evaluated using confusion matrices and different statistics such as the F1 score were reported. The study concludes with a recommendation for further development of datasets to train more accurate models without sacrificing real-time performance. The datasets should be generalized, including different lighting conditions and camera angles.