Enabling Robust Distracted Driving Performance Across Datasets with CLIP

Distracted driving is a major cause of road accidents. Detecting driver distraction in real-time using convolutional neural networks can significantly improve driving safety. However, existing methods relying on behavior classification to identify distraction suffer from high unreliability, especially in cross-dataset validations. This paper explores the use of CLIP, a recent contrastive language-image pre-training model, in distracted driving detection. Our cross-dataset tests on SFD, AUCDDV1, and 100-Driver suggest that CLIP holds promise in addressing the collapse issue observed in classification-based approaches across diverse datasets.