文献类型：期刊文献

英文题名：RT-DETR-DA for Complex Scenes: Distracted Driving Detection With Feature Interaction and Dynamic Perception

作者：Liu, Yi Li, Qiaoxing Wang, Xibin He, Ying Zhang, Sen

第一作者：刘毅;Liu, Yi

通信作者：Li, QX[1];Li, QX[2]

机构：[1]Guizhou Univ, Sch Management, Guiyang, Peoples R China;[2]Guizhou Univ, Collaborat Innovat Lab Digital Transformat & Gover, Guiyang, Peoples R China;[3]Key Lab New Power Syst Operat Control Guizhou Prov, Guiyang, Peoples R China;[4]Guizhou Inst Technol, Coll Big Data, Guiyang, Peoples R China;[5]Guizhou Univ Finance & Econ, Dept Acad Affairs, Guiyang, Peoples R China

第一机构：Guizhou Univ, Sch Management, Guiyang, Peoples R China

通信机构：corresponding author), Guizhou Univ, Sch Management, Guiyang, Peoples R China;corresponding author), Guizhou Univ, Collaborat Innovat Lab Digital Transformat & Gover, Guiyang, Peoples R China.

年份：2026

外文期刊名：ADVANCED INTELLIGENT SYSTEMS

收录：;EI(收录号：20261120274491);Scopus(收录号：2-s2.0-105032790264);WOS:【SCI-EXPANDED(收录号:WOS:001714531400001)】；

基金：The Project of Humanities and Social Sciences of Guizhou University (GDJD 202401).

语种：英文

外文关键词：complex scenes; cross-layer feature fusion; dual-path interaction; dynamic sparse gating mechanism; RT-DETR

摘要：Distracted driving is a leading cause of road accidents. This risk is particularly critical during autonomous driving system takeover requests, where a driver's inattention can lead to severe consequences. A core challenge for existing detection methods lies in their limited adaptability to real-world complex driving environments, such as variations across vehicle types and interference from multiple occupants. Specifically designed to address this issue, this paper proposes an enhanced detection framework based on real-time detection transformer. We boost model performance via two core modules: the dynamic sparse gating multiscale attention module, which strengthens multiscale feature extraction through a dynamic sparse gating mechanism, and the attention-guided dual-path fusion module, which achieves precise cross-layer feature fusion via dual-path interaction. Together, they significantly enhance the model's discriminative power and generalization capability in complex scenarios. Evaluation results on CBTDDD dataset demonstrate that the proposed method achieves a state-of-the-art balance between accuracy and speed, with 97.1% mAP50 and 63.6 FPS. This represents a 2.5% mAP50 improvement over the baseline model and outperforms other mainstream lightweight models. Visualization analyses further confirm its superior attention focus. This research provides an effective pathway for promoting the practical application of distracted driving detection.