面向ARM平台的自标定驾驶员疲劳检测方法

提出了一种面向进阶精简指令集机器(ARM)平台的自标定驾驶员疲劳检测方法。对驾驶员不同身高、体型及车内摄像头不同位置,采用驾驶员初始姿态自标定方法;采用改进的基于深度学习的多任务卷积神经网络(MTCNN),提取人脸识别和特征点,以得到头部姿态、眼睛、嘴巴运动等信息;基于操作员序列的深度卷积神经网络,来判断驾驶员的疲劳状态等级。实验了驾驶员疲劳检测方法。结果表明:相对于没有标定,采用本驾驶员自标定的方式,识别准确性提高了15%,采用MTCNN方法和ARM NEON加速技术,在“全志H5”、“树莓派”和Android手机上,运行速度分别是200、150、140 ms,提高约50%。因而,该检测方法,既提高了系统鲁棒性,也满足实时需求。

Self-calibration driver fatigue detection for advanced RISC machine(ARM)platform

A self-calibration driver fatigue detection method was proposed for Advanced RISC Machine(ARM)platform.Because of the difference of driver's height,body shape and the different installation position of camera,a driver's initial state self-calibration method was used to improve the system robustness.An improved multi task convolution neural network(MTCNN)based on deep learning was used to recognize the face and extract the feature points,to obtain the state information of head posture,eyes,mouth movement.A deep convolution neural network based on operator sequence was used to judge the driver's fatigue state level.The results show that the driver self-calibration method increases the recognition accuracy by 15%.Using MTCNN method and ARM neon acceleration technology increases the running speed of“Orange Pi Zero H5”,“Raspberry Pi”,and Android mobile phone by about 50%,respectively 200,150,and 140 ms.Therefore,this method not only improves the robustness of the system,but also meets the real-time requirements.