A Novel Method for Facial Expression-Based Driver Drowsiness Detection Leveraging Siamese Network in Adas Applications

Advanced Driver Assistance System (ADAS) is critical in enhancing road safety by providing real-time responses to drivers, especially in detecting abnormal behaviors, such as drowsiness. This paper proposed a drowsiness detection system that utilizes a Siamese Network based on facial expressions to accurately identify signs of deriver's fatigue. The Siamese Network employs two subnetworks, each trained to learn and differentiate between drowsy and non-drowsy state by analyzing facial expressions. The input data is presented as pairs, with each subnetwork processing one of the images in the pair to assess the difference between the two states. Several preprocessing techniques are applied to optimize the model's performance, including image rescaling and standardizing the input data for the network. During training process, Early Stopping is incorporated to prevent overfitting and ensure faster convergence of the model. Although Early Stopping contributed to improve generalization, it causes a slight reduction in accuracy. To mitigate this issue, the system was trained on multiple datasets and combinations of these datasets, which helped improve the overall performance. The resulting model achieves an accuracy of approximately 96 %, demonstrating its robustness, efficiency, and reliability for drowsiness detection in ADAS applications.