D4RNet: A lightweight distracted driver detection network with knowledge distillation and gradient reversal layer

Distracted Driver Detection (DDD) is critical for intelligent transportation systems, yet existing methods often incur high computational costs, leading to non-real-time detection on terminals and/or limited generalization across different camera views. In this work, we propose D4RNet , a lightweight D istracted D river D etection network that integrates knowledge D istillation with a gradient R eversal layer (GRL) for the first time in this domain. We employ a Swin-B as a teacher guide a MobileNetV2 student for training, in addition to predicting the type of distraction, the forward path extracts multi-scale, low-semantic features to predict the auxiliary tags, while during backpropagation the GRL reverses the gradient to adversarially suppress the auxiliary label, thereby enforcing the model to learn domain-invariant representations of the distraction behaviors, rather than noise such as image modality, camera angle, or vehicle scenes. Experimental results show improvements of 3.4% on D-all and 4.3% on N-all over the baseline, attains 81.32% average accuracy over eight subsets, surpassing the previous SOTA by 3.36%, while achieving 52 FPS on a real edge device, Jetson Nano via ONNX Runtime. Qualitative tests on real driving videos also demonstrate strong performance and robustness. Moreover, we introduce five new temporal–spatial combined splits for the 100-Driver dataset along with corresponding baselines, to provide a more comprehensive evaluation of model performance. We also indicate, based on numerous experiments, that future work should prioritize annotation quality and image diversity rather than architectural over-engineering. In summary, D4RNet offers a fast, accurate, and open-source solution for real-time DDD at the edge. Code, data, and models are available at https://github.com/jerodzhao/D4RNet .