Distraction Monitoring for Driving Safety: A Contrastive Learning Approach

The surge in road traffic accidents, largely attributable to driver distractions, necessitates advanced monitoring systems for vehicular safety. This paper introduces a novel network architecture grounded in contrastive learning to detect and assess the impact of driver distractions. The proposed system executes a two-stage task: classification of distracted behaviors and scoring of behavioral deviations from normal driving patterns. The network, which leverages a modified EfficientNet B4 as its backbone, demonstrates superior performance over traditional models like VGG16, ResNet50, and MobileNet v2. It achieves an accuracy of 96.25% and a macro F1 score of 96.13% on the public dataset with relatively low computational cost, showcasing its effectiveness in real-world applications. The study also explores the use of different encoders and highlights the network’s interpretability and adaptability, aligning with the growing needs of the automotive industry for Driver Monitoring Systems (DMS) and Advanced Driver-Assistance Systems (ADAS). The paper concludes with a discussion on the network’s practical deployment, promising future integration into vehicles for enhanced road safety.