An Efficient Deep Learning Technique for Driver Drowsiness Detection

Deep learning techniques allow us to learn about a person’s behavior based on pictures and videos. Using digital cameras, the system can identify and classify a person’s behavior based on images and videos. This paper aims to present a method for detecting drivers’ drowsiness based on deep learning. To determine which transfer learning technique best suits this work, we used DenseNet169, MobileNetV2, ResNet50V2, VGG19, InceptionV3, and Xception on the dataset. The dataset used in this paper is the Driver Drowsiness Dataset (DDD), which is publicly available on Kaggle. This dataset consists of 41,790 RGB images, and each image has a size of 227 $$\times$$ 227, which has 2 classes: drowsy and not drowsy. The Drivers Drowsiness Dataset is based on the images extracted from the real-life Drowsiness dataset (RLDD). After comparing the results coming from all 6 models, the highest accuracy achieved was 100% by ResNet50V2, and various parameters are calculated like accuracy, F1 score, etc. Additionally, this work compared the results with existing methods to demonstrate its effectiveness.