| 摘要: |
| 目的 驾驶员疲劳是导致交通事故的主要因素之一,尤其在夜间低光照环境下,疲劳检测变得更加具有挑战
性。 针对低光环境下驾驶员面部特征提取困难的问题,提出了一种融合局部全局特征的低光照环境驾驶员疲劳检
测方法。 方法 该方 法 包 括 低 光 增 强 模 块 和 双 流 网 络 检 测 模 块。 低 光 增 强 模 块 是 在 多 分 枝 弱 光 增 强 网 络
( MBLLEN) 的特征提取层添加通道注意力( SE-Net) 模块,提高低光增强算法对驾驶员面部特征信息的关注,更好
地增强了夜间驾驶员的低光图像。 双流网络检测模块使用双分支网络结构,分别提取全局面部信息和局部面部信
息。 首先,驾驶员原始的夜间图像经过低光增强算法增强; 然后, 输入双流网络中, 双分支结构网络分别使用
ResNet-34 和 ResNet-18 提取全局特征和局部特征;最后,通过集成学习方法将局部和全局特征信息融合,以不同
权重贡献比的方式生成最终的疲劳状态预测。 结果 实验结果表明:提出的方法在 NTHU-DDD 数据集上表现出良
好的性能,最终的检测准确率为 90. 10%。 结论 文中提出的方法在夜间驾驶员疲劳检测方面表现出较高的准确性,
为夜间驾驶员疲劳检测领域带来了重要的进展,具有深远而重要的意义。 |
| 关键词: 疲劳检测 低光增强 双分支网络 集成学习 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Driver Fatigue Detection Method in Low-light Environments |
|
YAO Wei
|
|
School of Computer Science and Engineering Anhui University of Science and Technology Anhui Huainan 232001
China
|
| Abstract: |
| Objective Driver fatigue is a major factor contributing to traffic accidents especially in low-light environments
at night where fatigue detection becomes more challenging. To address the difficulty in extracting the facial features of
drivers under low-light conditions a method combining local and global features was proposed for driver fatigue detection
in low-light environments. Methods The proposed approach comprised a low-light enhancement module and a dual-stream
detection module. The low-light enhancement module incorporated a channel attention mechanism SE-Net into the
feature extraction layer of the multi-branch low-light enhancement network MBLLEN . This enhanced the algorithm?? s
focus on the driver?? s facial feature information improving the quality of low-light images of drivers at night. The dual-
stream detection module employed a dual-branch network structure to extract global and local facial features. First the
original nighttime images of drivers were enhanced using the low-light enhancement algorithm. These enhanced images
were then input into the dual-stream network where the two-branch structured network used ResNet-34 and ResNet-18 to
extract global and local features respectively. Finally the local and global feature information was fused using an
ensemble learning method and the final fatigue state prediction was generated with different weight contribution ratios.
Results The experimental results show that the proposed method exhibited good performance on the NTHU-DDD dataset, and the final detection accuracy was 90. 10%. Conclusion The method proposed in this paper shows high accuracy in
nighttime driver fatigue detection and brings important progress in the field of nighttime driver fatigue detection which is
of far-reaching and important significance. |
| Key words: fatigue detection low-light enhancement dual-branch network ensemble learning |
