| 摘 要: 为解决夜间低光环境下驾驶员困倦检测面临的图像特征模糊、小目标识别困难及实时性能不足等问题,本文基于改进的YOLOv11-n架构提出一种轻量级检测方法——LLE-YOLO(Low-Light Enhancement YOLO)。首先,在主干网络中引用CondConv模块,增强低光环境下眼部纹理特征的表征能力。其次,使用 ODConv模块重构颈部网络,降低冗余性并提升推理效率。最后,在头部特征层嵌入Swin Transformer模块,增强遮挡场景中小目标特征融合能力。基于自建夜间场景数据集的实验结果表明,LLE-YOLO模型达到91.5%的平均精确率(mAP)和1%的假阳性率(FPR),为智能驾驶舱环境下的驾驶员状态监测系统提供了稳健且工程可行的解决方案。 |
| 关键词: 驾驶员疲劳检测 YOLOv11 CondConv ODConv Swin Transformer |
|
中图分类号:
文献标识码:
|
|
| Low-Light Driver Drowsiness Detection Algorithm Improved Based on YOLOv11ZHANG Lanxiang, Lv Hongming, Jing Fusheng, Luo Junfeng |
|
zhanglanxiang
|
Yancheng Institute of Technology, School of Automotive Engineering
|
| Abstract: To address challenges of obscured image features, small-target recognition difficulties, and inadequate real-time performance in driver drowsiness detection in nighttime low-light conditions, this paper proposes a lightweight detection method termed LLE-YOLO (Low-Light Enhancement YOLO) based on an improved YOLOv11-n architecture. First, the CondConv module is incorporated into the backbone network to enhance the representation of eye texture features in low-light conditions. Second, the ODConv module is used to reconstruct the neck network, reducing redundancy and improving inference efficiency. Finally, the Swin Transformer module is embedded in the head feature layer to enhance small object feature fusion in occlusion scenarios. Experimental results based on a self-built nighttime scene dataset demonstrate that the LLE-YOLO model achieves a mean average precision (mAP) of 91.5% and a false positive rate (FPR) of 1%, providing a robust and engineering-feasible solution for driver state monitoring systems in intelligent cockpit environments. |
| Keywords: Driver drowsiness detection, YOLOv11, CondConv, ODConv, Swin Transformer |
