DrowsyDetectNet: Driver Drowsiness Detection Using Lightweight CNN With Limited Training Data

Abstract

One major factor in the rising incidence of traffic accidents is the driver’s drowsiness. Innovations in computer vision technology have made it possible to construct smart cams that can recognize driver fatigue. By alerting drivers, this technology successfully lowers the total number of accidents caused by weariness. This study proposes a DrowsyDetectNet that utilizes a shallow Convolutional Neural Network (CNN) architecture to identify driver drowsiness. The 68-point face landmark identification approach is used to identify faces and extract eye areas. The proposed system employs a shallow CNN architecture with fewer layers and parameters to detect driver drowsiness with limited training data. Feature extraction focuses on relevant visual cues for drowsiness detection, such as eyelid closure. The transfer learning models, such as VGG19, ResNet50, MobileNetV2, and InceptionV3, are also used to identify driver drowsiness. Two datasets, Dataset-1 and Dataset-2, were utilized to assess this study. On two datasets, the proposed DrowsyDetectNet produced an accuracy of 99.23% and 99.14%, respectively. The proposed DrowsyDetctNet framework achieved better accuracy when compared with state-of-the-art models and pre-trained models.

Keywords

• Drowsiness detection
• DrowsyDetectNet
• shallow CNN
• limited training data
• pre-trained models