| 基于防疫常态化的驾驶员疲劳状态检测方法 |
| |
| 引用本文： | 黄玲,洪佩鑫,吴泽荣,刘建荣,黄子虚,崔躜.基于防疫常态化的驾驶员疲劳状态检测方法[J].交通信息与安全,2021,39(4):26-34. |
| |
| 作者姓名： | 黄玲 洪佩鑫 吴泽荣 刘建荣 黄子虚 崔躜 |
| |
| 作者单位： | 华南理工大学土木与交通学院 广州 510640 |
| |
| 基金项目： | 国家自然科学基金项目51775565广州市重点区域研发计划项目202007050004广东省教育厅项目2020KQNCX205 |
| |
| 摘 要： | 疲劳驾驶检测是交通安全领域的研究分支, 而新冠疫情形势下口罩的佩戴又提出了新的挑战。为此通过基于ResNet-10的SSD模型检测驾驶员人脸, 并使用MobileNet-V2轻量级模型判断是否佩戴口罩, 测试集验证该分类器可以达到98.50%的判断精度。在未佩戴口罩的情况下采用传统图像HOG特征结合SVM分类器检测驾驶员人脸。在后续处理中利用级联回归器定位特征点和提取时间窗口内的疲劳指标, 采用二次判定对疲劳状态采取文字和声音预警, 而在清醒状态下会调整各项判断阈值。对算法在预采集的视频样本和NTHU-DDD测试集下进行测试, 验证了该框架能以18.42帧/s的总体速度实现92.65%和86.09%的检测精度。实验结果表明, 该框架应对佩戴眼镜、脸部姿态变化和光照条件差异具有强鲁棒性, 而且能够兼顾疲劳检测的口罩干扰和实时性。
|
| 关 键 词： | 智能交通 疲劳检测 自适应判断 口罩分类 |
| 收稿时间： | 2021-06-02 |
A Detection Method for Drivers' Fatigue States Based on Normalization of Epidemic Prevention |
| |
| HUANG Ling,HONG Peixin,WU Zerong,LIU Jianrong,HUANG Zixu,CUI Zuan.A Detection Method for Drivers' Fatigue States Based on Normalization of Epidemic Prevention[J].Journal of Transport Information and Safety,2021,39(4):26-34. |
| |
| Authors: | HUANG Ling HONG Peixin WU Zerong LIU Jianrong HUANG Zixu CUI Zuan |
| |
| Affiliation: | School of Civil Engineering and Transportation, South China University of Technology, Guangzhou 510640, China |
| |
| Abstract: | The detection of fatigue driving is a research branch of traffic safety, and wearing masks in the COVID-19 situation poses a new challenge. Therefore, the driver's face is detected by the single-shot multi-box detector(SSD)model based on ResNet-10, and the MobileNet-V2 model is used to classify masks. The test set verifies that the classifier can reach an accuracy of 98.50%. The histogram of the oriented gradient(HOG)feature combined with the support vector machine(SVM)classifier is used to detect the driver's face without wearing a mask. In the subsequent processing, the cascade regress is used to locate the feature points and extract the fatigue indices in the time window. The second judgment is used to perform the text and sound warnings for the fatigue state, and the judgment thresholds are adjusted in the awaken state. The algorithm experimented on pre-collected videos and NTHU-DDD can achieve the accuracy of 92.65 and 86.09% at the overall speed of 18.42 fps, respectively. The proposed framework shows strong robustness against the variation of wearing glasses, facial posture, and illumination, considering the interference of mask and real-time performance. |
| |
| Keywords: | |
| 本文献已被 CNKI 万方数据 等数据库收录！ |
| 点击此处可从《交通信息与安全》浏览原始摘要信息 |
|
点击此处可从《交通信息与安全》下载全文 |
|
