基于级联卷积神经网络的疲劳检测

发布时间：2018-11-10 11:46

【摘要】：基于计算机视觉的疲劳检测具有低侵入性、低成本的优点,然而光照变化、面部表情、复杂背景等仍然对检测率造成很大的影响。以卷积神经网络为代表的深度学习以其较强的特征提取能力和鲁棒性在模式识别领域取得了成功的应用。本文提出了一种基于级联卷积神经网络(CNN)结构的疲劳检测算法。首先训练第1级网络实现人眼与非人眼的分类,使网络充分学习人眼特征,当输入目标图像时,人眼区域能快速从第一级网络特征图中分离出来;然后将人眼图像传送给第2级网络检测眼部特征点位置,计算眼睛张开度并以此判断测试者眼睛状态,构造疲劳检测模型;最后根据连续多帧的眼睛状态序列,判断测试者是否处于疲劳状态。在检测误差为5%时,眼部4个特征点的平均检测正确率为93.10%,单点检测正确率最高可达97.14%。测试结果表明,在本文提出方法下眼睛的清醒和疲劳状态有明显的不同,证明本文提出的方法有效可行,具有较好的应用前景。
[Abstract]:Fatigue detection based on computer vision has the advantages of low invasion and low cost. However, illumination changes, facial expressions and complex background still have a great impact on the detection rate. The deep learning represented by convolution neural network has been successfully applied in pattern recognition with its strong feature extraction ability and robustness. This paper presents a fatigue detection algorithm based on cascaded convolution neural network (CNN) structure. First, the first level network is trained to realize the classification of human eye and non-human eye, so that the network can fully learn the human eye feature. When the target image is inputted, the human eye region can be quickly separated from the first level network feature map. Then the human eye image is transmitted to the second level network to detect the position of eye feature points, and the eye opening degree is calculated to judge the state of the testers' eyes, and a fatigue detection model is constructed. Finally, according to the sequence of eye states, the testers are judged whether they are tired or not. When the detection error is 5, the average detection accuracy of the four characteristic points of the eye is 93.1010 and the highest accuracy of single point detection is 97.14. The test results show that there are obvious differences between the awake and fatigue states of the eyes under the proposed method, which proves that the proposed method is effective and feasible, and has a good prospect of application.
【作者单位】：南开大学现代光学研究所;公安海警学院电子技术系;
【基金】：国家自然科学基金(61401105) 教育部重点(DIA150308) 浙江省教育教学改革(JG2015207)资助项目
【分类号】：TP183;TP391.41

【相似文献】