DRDG-Net: Generalizable Driver Distraction Detection with Decoupled Representation and Gradient Reversal Layer

Driver Behavior Recognition (DBR) is crucial for traffic safety and intelligent transportation. However, domain shift and individual behavior differences still hinder generalization across drivers and datasets. To address this, we propose DRDG-Net(Decoupled Representation with Domain Generalization Network), a novel multimodal framework that combines visual input and identity indices to learn behavior-related, identity-independent features. It incorporates behavior-identity decoupling, semantic-guided fusion for cross-modal alignment, adversarial domain learning for robustness, fuzzy classification (Score-Softmax) to handle ambiguous boundaries, and learnable loss-weighting for task balance. Experiments on three public datasets—StateFarm, AUC-V1, and 100-DriverMini—show that DRDG-Net consistently outperforms baselines under crossdomain conditions, demonstrating strong robustness and generalization.