Keeping Drivers Focused: A Deep Learning Model for Driver Distraction Detection

Driver distraction detection continues to be a highly researched topic. Driver monitoring systems provide a twofold benefit of alerting drivers and at the same time enabling fleet management companies to verify the reliability of their drivers based on collected data. In this paper, we present a novel approach to improving driver distraction detection systems by using a “neural style transfer” based augmentation method and multi-view learning techniques. Our primary contributions include the nstAugment function, which provides an easily customizable method for performing “neural style transfer” based augmentation, and a “meta-learner layer” for effective multi-view integration. The nstAugment function significantly enhances model robustness, providing an accuracy improvement of 7% on side-view model and 2% on multi-view model. Additionally, the integration of the “meta learner layer” facilitates comprehensive analysis from multiple perspectives, achieving state-of-the-art (SOTA) performance on the 3MDAD dataset with an accuracy of 87.44%. The results show the effectiveness of the meta-learner as well as the nstAugment function in improving model accuracy.