Scholarlink

<p>Driver state recognition is a critical component of intelligent transportation systems, facilitating early detection of distraction, drowsiness, and unsafe behaviors that significantly contribute to road accidents. This work presents a unified comparative study of hybrid deep learning architectures for fine-grained multi-class driver state recognition, concentrating on the complementary roles of global attention and spatial relational reasoning. A nine-class driver behavior dataset was constructed by merging and re-annotating three public sources, followed by standardized preprocessing and class-weighted training to address real-world data imbalance. Within a consistent CNN-Transformer framework, multiple attention mechanisms, SE, CBAM, CSA, and Efficient Channel Attention (ECA), are comprehensively evaluated against a CNN + Graph Neural Network (GNN) baseline that explicitly models spatial relationships between driver regions. Experimental results on a held-out test set of 3,103 images demonstrate that while the CNN + GNN model achieves strong overall performance (accuracy ≈ 99%), it demonstrates reduced sensitivity in fine-grained classes involving subtle eye and facial cues. In contrast, Transformer-based models consistently improve validation accuracy (99.35%–99.48%), with the proposed CNN-Transformer + ECA obtaining the best balance of accuracy, stability, and generalization, reaching near-perfect test performance and AUC values approaching 1.00 across all classes. Analysis of confusion matrices, training dynamics, and ROC curves confirms that lightweight channel-wise attention combined with global self-attention enhances discriminative capability without increasing architectural complexity. These findings highlight the practical efficacy of unified CNN-Transformer designs with efficient attention for reliable, real-time driver monitoring applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cross-architecture evaluation for multi-class driver state recognition: unveiling the performance gap between hybrid transformers and gnns

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Md. Ariful Islam,
  • Md Al Emran,
  • Md. Wajed Ali

摘要

Driver state recognition is a critical component of intelligent transportation systems, facilitating early detection of distraction, drowsiness, and unsafe behaviors that significantly contribute to road accidents. This work presents a unified comparative study of hybrid deep learning architectures for fine-grained multi-class driver state recognition, concentrating on the complementary roles of global attention and spatial relational reasoning. A nine-class driver behavior dataset was constructed by merging and re-annotating three public sources, followed by standardized preprocessing and class-weighted training to address real-world data imbalance. Within a consistent CNN-Transformer framework, multiple attention mechanisms, SE, CBAM, CSA, and Efficient Channel Attention (ECA), are comprehensively evaluated against a CNN + Graph Neural Network (GNN) baseline that explicitly models spatial relationships between driver regions. Experimental results on a held-out test set of 3,103 images demonstrate that while the CNN + GNN model achieves strong overall performance (accuracy ≈ 99%), it demonstrates reduced sensitivity in fine-grained classes involving subtle eye and facial cues. In contrast, Transformer-based models consistently improve validation accuracy (99.35%–99.48%), with the proposed CNN-Transformer + ECA obtaining the best balance of accuracy, stability, and generalization, reaching near-perfect test performance and AUC values approaching 1.00 across all classes. Analysis of confusion matrices, training dynamics, and ROC curves confirms that lightweight channel-wise attention combined with global self-attention enhances discriminative capability without increasing architectural complexity. These findings highlight the practical efficacy of unified CNN-Transformer designs with efficient attention for reliable, real-time driver monitoring applications.