基于神经网络的疲劳驾驶检测方法研究

Research on Fatigue Driving Detection Method Based on Neural Network

摘要: 由于疲劳驾驶是导致交通事故的重要原因。因此，针对疲劳驾驶检测，本文提出一种基于神经网络的疲劳检测方法。对于传统疲劳检测技术检测精度不高、综合检测能力不足等问题提出了相应的改进策略，这对预防疲劳驾驶、提高驾驶安全具有重要的意义。本文的主要工作如下：1) 采用基于Retina Face算法的人脸检测网络，同时本文收集了大量局部遮挡和大姿态偏转的人脸图像作为训练样本，使得人脸检测模型更加灵活，解决了传统人脸检测算法对于人脸姿态要求过严的缺点。2) 本文从驾驶员眼部、嘴部以及头部姿态出发，基于神经网络的方法，综合分析驾驶员的疲劳状态。当系统存在某一指标失效时，依然可以正常检测出驾驶员的疲劳状态，具有一定的容错性。此外，借鉴PERCLOS经验准则，提出了基于一定时间周期的疲劳判定法，使系统对于驾驶员的疲劳预估更加准确。实验结果表明，使用本文提出的方法，能够实现多特征点的定位与识别，且基于眼部状态以及嘴部状态预测疲劳的精度分别达到了97.1%和97.5%，根据头部状态预测疲劳的精度达到了88.1%，处理速度也达到了每张图片62 ms。相较于传统只基于眼部状态检测疲劳的方法，本文提出的算法在保证准确率和效率的前提下，提高了综合检测能力。

Abstract: Fatigue driving is an important cause of traffic accidents. Therefore, for fatigue driving detection, this paper puts forward a fatigue detection technology based on deep learning, and puts forward corresponding improvement strategies for the problems of low detection accuracy and poor real-time performance of traditional fatigue detection technology, which is of great significance to prevent fatigue driving and improve driving safety. The main work of this paper is as follows: 1) The face detection network based on Retina Face algorithm is adopted, and a large number of face images with partial occlusion and large posture deflection are collected as training samples, which makes the face detection model more flexible and solves the shortcoming that the traditional face detection algorithm requires too strict face posture. 2) Based on the attitude of the driver’s eyes, mouth and head, this paper comprehensively analyzes the driver’s fatigue state based on the method of neural network. When a certain index of the system fails, the fatigue state of the driver can still be detected normally, which has a high enough fault tolerance rate. In addition, a fatigue judgment method based on a certain time period is proposed by referring to PERCLOS, which makes the system more accurate for driver fatigue prediction. The experimental results show that the method proposed in this paper can realize the location and recognition of multi-feature points, and the accuracy of fatigue prediction base on the eye state and mouth state is 97.1% and 97.5% respectively. The accuracy of fatigue prediction base on the head state is 88.1%, and the processing speed of each image is 62 ms. Compared with the traditional fatigue detection methods of only de-tecting eye state, the algorithm proposed in this paper improves the comprehensive detection ca-pability, based on sufficient accuracy and efficiency.