摘要： 现有疲劳驾驶检测方法通常将驾驶过程中采集的数据传输至云端进行分析，然而在车辆移动过程中网络覆盖范围、响应速度等因素会造成检测实时性差。为在车载嵌入式设备上对驾驶人疲劳状态进行准确预警，提出一种基于边缘计算的疲劳驾驶检测方法。通过改进的多任务卷积神经网络确定人脸区域，根据人脸的面部比例关系定位驾驶人的眼部与嘴部区域，利用基于Ghost模块的轻量化AlexNet分类检测眼部与嘴部的开闭状态，并结合PERCLOS和PMOT指标值实现疲劳检测。在NHTU-DDD数据集上的实验结果表明，该方法在树莓派4B开发板上的检测准确率达到93.5%且单帧平均检测时间为180 ms，在保障检测准确率的同时大幅降低了计算量，能较好地满足疲劳驾驶的实时检测需求。

关键词: 疲劳驾驶检测, 边缘计算, 多任务卷积神经网络, 轻量化, AlexNet结构

Abstract: Most of the existing fatigue driving detection methods require the collected driving data to be sent to the cloud for analysis.However,the real-time performance of the methods is often affected by the network coverage,response speed and other factors as the vehicle moves.To improve the accuracy of fatigue alarms generated on onboard embedded devices,a fatigue driving detection method based on edge computing is proposed.The method employs an improved Multi-Task Convolutional Neural Network(MTCNN) to determine the facial area,and then the parts of the eyes and the mouth based on facial proportions.The lightweight AlexNet using the Ghost module is used to detect whether the eyes and the mouth are closed.On this basis,the fatigue driving detection is realized with the indicators of PERCLOS and PMOT considered,too.Experimental results on the NHTU-DDD dataset show that the method achieves an accuracy of 93.5% on Raspberry Pi 4B,and its average time for single-frame prediction is about 180 ms.The method ensures the detection accuracy while significantly reducing the amount of calculation,meeting the demand for real-time fatigue detection.

Key words: fatigue driving detection, edge computing, Multi-Task Convolutional Neural Network(MTCNN), lightweight, AlexNet structure

中图分类号: 

• TP391