MGS-DETR: a multi-granularity spatial representation learning framework for joint driver fatigue and distraction detection

Driver fatigue and distracted driving are major causes of traffic accidents, yet existing vision-based detectors often suffer from imbalanced feature modeling, insufficient cross-layer fusion, and loss of fine-grained spatial details under complex in-vehicle conditions. To address these issues, we propose MGS-DETR, a multi-granularity spatial representation learning framework built upon RT-DETR for joint driver fatigue and distraction detection. In the encoding stage, we develop a Multi-scale Spatial Awareness Encoder (MSAE) to balance local fine-grained fatigue cues and global distracted-action context.

In the feature-fusion stage, we design Global-guided Cross-Modulation Fusion (G-CMF) to enhance the interaction between shallow spatial details and deep semantic features under pose variation and occlusion. In the multi-scale transmission stage, we construct a Spatial-Detail Enhanced Feature Pyramid Network (SDE-FPN) to preserve fine-grained spatial details during feature transmission and improve the localization of annotated regions for fatigue-related visual signs. Experiments are conducted on the Composite Distraction and Fatigue Dataset (CDFD) and the self-collected Real-World Fatigue and Distracted Driving Dataset (RWFDD).

On CDFD, MGS-DETR achieves 0.9628 precision, 0.9618 recall, 0.9708 mAP50, and 0.8159 mAP50-95, improving mAP50-95 by 5.44 percentage points over RT-DETR-r18. Under the source-video-group-disjoint RWFDD split, it obtains 0.8868 precision, 0.8946 recall, 0.8907 F1-score, 0.9123 mAP50, and 0.8501 mAP50-95. Bidirectional transfer experiments are reported separately under the stated nine-class protocol.

The full model runs at 164.50 FPS with 21.90 M parameters and 64.0 GFLOPs on the desktop test platform; embedded-device throughput is measured separately on Jetson Orin NX. The authors thank all drivers who agreed to participate in the data acquisition work. This work was supported by the Innovative Entrepreneurship Undergraduate Training Programme (Grant No.

Yz2025437) at Yangtze University. The authors gratefully acknowledge this support. Lingjian Kong and Yongli Zhang contributed equally to this work.

School of Electronic Information and Electrical Engineering, Yangtze University, Jingzhou, 434023, Hubei, China Lingjian Kong, Yongli Zhang, Ziyu Qiu, Hongying Qin & Kai Xie School of Future Technology, Yangtze University, Jingzhou, 434023, Hubei, China School of Computer Science, Yangtze University, Jingzhou, 434023, Hubei, China School of Telecommunications Engineering, Xidian University, Xi’an, 710068, Shaanxi, China School of Electronic Information, Central South University, Changsha, 410083, Hunan, China The authors declare no competing interests. Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material.