| 摘 要: 针对传统依赖人脸关键点的驾驶员面部特征检测方法存在耗时、稳定性差的问题,本文提出一种轻量化的驾驶员面部特征检测算-法YOLOv8_FDD(YOLOv8-based Fatigue Drowsiness Detection),该算法是基于YOLOv8架构进行改进的。首先,通过引入SimAM(Simple Parameter-Free Attention Module)注意力机制,有效提高了模型对小目标的检测精度和查全率。其次,在路径聚合网络的基础上,新增了两条跨尺度连接路径,增强了模型的特征融合能力,减少了模型参数量。再次,通过引入PConv(Partial Convolution)设计了轻量化检测头,进一步提高了模型的检测精度,降低了模型的复杂度。最后,引入 WIoUv3(Wise-IoU version3)损失函数,提高了模型对小目标的定位和检测能力。在疲劳驾驶数据集 FDD(Fatigue Driving Dataset)上的实验结果显示,本文提出YOLOv8_FDD(YOLOv8-based Fatigue Drowsiness Detection)模型平均检测精度均值为95.9%,检测速度为161.2fps,其性能优于 YOLOv8n算法的性能,非常适用于疲劳驾驶的实时监测场景。 |
| 关键词: 驾驶员面部特征检测;YOLOv8;SimAM 注意力机制;PConv;WIoUv3 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 陕西省自然科学基金项目(2024JC-YBQN-0724) |
|
| Research on DriverFacial Feature Detection Algorithm Based on YOLOv8_FDD |
|
LIANG Yanrong, WANG Xiaoxia, CHEN Xiao
|
(School of Electronic Information and Artificial Intelligence, Shaanxi University of Science & Technology, X'i an 710021, China)
18049411296@163.com; wangxiaoxia@sust.edu.cn;; chenxiao@sust.edu.cn
|
| Abstract: To address the time-consuming and unstable issues of the traditional driver facial feature detection methods that rely on facial key points, this paper proposes YOLOv8 _ FDD (YOLOv8-based Fatigue Drowsiness Detection), a lightweight driver facial feature detection algorithm based on an improved YOLOv8 architecture. Firstly, the SimAM ( Simple Paramete-r Free Attention Module) attention mechanism is introduced to enhance detection accuracy and recall rates for small targets. Secondly, two additional cross-scale connection paths are integrated into the Path Aggregation Network (PAN) to strengthen feature fusion capability while reducing model parameters. Thirdly, a lightweight detection head is designed using PConv (Partial Convolution) to further improve detection accuracy and reduce computational complexity. Finally, the WIoUv3 (Wise-IoU version 3) loss function is adopted to enhance localization and detection performance for small targets. Experimental results on the Fatigue Driving Dataset (FDD) demonstrate that the proposed YOLOv8_FDD achieves a mean Average Precision (mAP@ 0.5) of 95.9% and an inference speed of 161.2 fps, outperforming the baseline YOLOv8n algorithm and proving highly suitable for rea-l time monitoring of fatigue driving. |
| Keywords: driver facial feature detection; YOLOv8; SimAM attention mechanism; PConv; WIoUv3 |
