Self-Supervised Learning for Driver Distraction Detection: A Comparative Study with Traditional Supervised Models

Abstract

Distracted driving accounts for a significant proportion of global road fatalities, yet current detection systems require prohibitively expensive labeled datasets. We introduce a Self-Supervised Learning (SSL) framework that fundamentally transforms the data efficiency paradigm for driver distraction detection. By leveraging Bootstrap Your Own Latent (BYOL) contrastive learning on unlabeled driver images, our ResNet50- based architecture achieves 97.37% classification accuracy on the State Farm benchmark while requiring labeled supervision only during a lightweight linear probing phase. This represents merely a 2% accuracy trade-off compared to the best fully supervised method (99.32%), yet eliminates the need for extensive manual annotation during feature learning. Our approach demonstrates strong cross-driver generalization through driver-disjoint data splitting, handling challenging intra-class variations including illumination changes, pose differences, and camera angles. Systematic ablation studies quantify the contribution of each architectural component, while Grad-CAM visualizations provide interpretable evidence of learned attention mechanisms. To our knowledge, this is the first comprehensive application of modern SSL techniques to the State Farm dataset, establishing a new paradigm for data-efficient driver monitoring that is particularly valuable for real-world deployments where annotation resources are constrained and rapid adaptation to new drivers is essential.