ENHANCING MODEL PERFORMANCE OF AUTOMATIC DRIVER DISTRACTION DETECTION USING TRANSFER LEARNING

Abstract

Video-based anomalous driving behavior identification is becoming more and more important for the sake of driver and passenger safety and  the development of autonomous driving . Thanks to recent developments in deep learning, which make use of the generalizability of complex models and training on massive video datasets, this work has become much easier. Various deep learning algorithms are investigated in this work, such as VGG16, MobileNetV2, DenseNet, and a fusion model that combines VGG16 and DenseNet. After extensive testing, DenseNet was found to be the most accurate, with a score of 0.883. This shows how well it can identify suspicious driving habits. In order to tackle this difficult detection problem, DenseNet is crucial, thanks to its dense connections that improve gradient flow. In order to speed up the development of autonomous driving technology, these results demonstrate the promise of deep learning methods, and DenseNet in particular, for improving highway safety.

Densely linked convolutional network (DenseNet), VGG16, MobileNetV2, Automatic driving, Safety Accuracy, and video-based aberrant driving behavior identification are some of the keywords associated with deep learning methods.