《计算机应用研究》|Application Research of Computers

基于聚类框架与局部感受野的实时人脸疲劳检测

Real-time face fatigue detection based on receptive field and clustering

免费全文下载 (已被下载 次)  
获取PDF全文
作者 刘君扬,王金凤
机构 华南农业大学 数学与信息学院,广州 510642
统计 摘要被查看 次,已被下载
文章编号 1001-3695(2020)12-057-3795-04
DOI 10.19734/j.issn.1001-3695.2019.07.0315
摘要 针对在自然环境下人脸疲劳识别遇到的问题,如人脸检测率不高、判别疲劳的特征过于单一、检测速度慢等,提出了一种基于聚类框架与局部感受野的实时人脸疲劳检测方法。首先对人脸尺寸进行聚类分析,根据聚类类别决定检测层个数并设置先验框大小,根据预测特征图的感受野与人脸尺寸匹配的原则设置网络层数,最后通过最小化损失函数学习多种疲劳特征。实验证明,在驾驶室等环境下基于聚类框架与局部感受野的方法在保持识别准确率的同时提高了检测速度,使用GPU GeForce GTX TITAN能达到125 fps,满足了实时性要求。
关键词 神经网络; 深度学习; 目标检测; 疲劳识别; 感受野; 聚类
基金项目 广东省科技计划资助项目(2017A040406023)
广州市科技计划资助项目(201804010353)
本文URL http://www.arocmag.com/article/01-2020-12-057.html
英文标题 Real-time face fatigue detection based on receptive field and clustering
作者英文名 Liu Junyang, Wang Jinfeng
机构英文名 College of Mathematics & Informatics,South China Agricultural University,Guangzhou 510642,China
英文摘要 Faced with the problems encountered in fatigue detection under natural environment, such as low detection rate of the face, slow detection speed and single feature for judging, etc., this paper proposed a fatigue detection method combining receptive field with clustering algorithm. Firstly, it applied cluster on the face size and returned the cluster numbers which determined the number of detection layer. Then, it set the size of anchor boxes according to the face size. Secondly, the algorithm set the number of convolutional layers according to the principle that the receptive field should match the face size in the predicted feature map. Finally, new algorithm learnt a variety of fatigue characteristics by minimizing the loss function. Experiments show that this method based on clustering framework and local receptive field have improved the detection speed while maintaining the recognition accuracy. It can reach 125 fps by using GPU GeForce GTX TITAN, and satisfy the request of real time.
英文关键词 neural network; deep learning; object detection; fatigue identification; receptive field; cluster
参考文献 查看稿件参考文献
 
收稿日期 2019/7/19
修回日期 2019/9/8
页码 3795-3798
中图分类号 TP391.4
文献标志码 A