<p>随着经济的发展，全世界每年的汽车保有量都在飞速上升，汽车数量的增多势必带来交通事故的频发。而在这些交通事故中，有1成以上是由于驾驶员疲劳驾驶所引起的。因此，实现疲劳驾驶的检测对保护驾驶员生命和财产安全有重要意义。</p> <p>本文的目的是研究设计一套基于面部特征的疲劳驾驶检测系统，该系统能够基于目标检测算法实现疲劳状态判定，同时加入疲劳延时判断功能，当系统判定驾驶员处于疲劳驾驶时会发出语音报警信息，确保驾驶员意识清醒，从而避免事故的发生。本文的主要内容包括面部目标测研究、模型压缩研究和系统实现及测试。</p> <p>本文使用目标检测经典网络YOLOv4框架作为目标训练模型，针对眼部和嘴部疲劳特征提取问题，本文对目标检测模型进行改进：在原始模型的基础上加入空洞卷积和局部卷积层，加深了特征的提取，提高了感受野范围。改进后的YOLOv4算法mAP值达到97.29%，相较原YOLOv4mAP值提高了1.98%，其中对眼睛部位的检测提高了6%。在疲劳判定方面，针对疲劳检测实时性的问题，本文通过疲劳动作的持续时间进行判断。模型检测时首先检测到人脸，然后对眼部和嘴部特征进行检测，如果检测到疲劳特征，进行画框处理，如果疲劳特征持续达到2秒，系统会进行疲劳报警提醒。除此以外，为了更好的适配不同性能的硬件设备，在模型部署之前使用了模型压缩技术对模型进行压缩加速，通过对比模型剪枝、模型量化和TensorRT模型压缩三种方法对本模型的实际效果，选取TensorRT模型加速方法，该方法使得模型权重大小缩小一倍，使模型能够部署在内存更小的设备上。在本机加速后帧率由原来的18变为40，部署在硬件上的帧率由3变成9，满足疲劳监测的实时性和准确性。</p>

<p>With the development of the economy, the annual car ownership in the world is increasing rapidly, and the increase in the number of cars will inevitably lead to frequent traffic accidents. In these traffic accidents, more than 10% are caused by driver fatigue driving. Therefore, it is of great significance to realize the detection of fatigue driving to protect the driver&#39;s life and property safety.</p> <p>The purpose of this thesis&nbsp;is to study and design a fatigue driving detection system based on facial features. The system can realize the fatigue state judgment based on the target detection algorithm, and at the same time add the fatigue delay judgment function. When the system judges that the driver is driving fatigued, it will send out a voice alarm. information to ensure that the driver is conscious and avoid accidents. The main content of this thesis&nbsp;includes the research of facial target measurement, model compression research and system implementation and testing.</p> <p>This thesis&nbsp;uses the classic target detection network YOLOv4 framework as the target training model. Aiming at the problem of eye and mouth fatigue feature extraction, this thesis&nbsp;improves the target detection model: adding dilated convolution and local convolution layers on the basis of the original model, deepening the The extraction of features improves the scope of the receptive field. The mAP value of the improved YOLOv4 algorithm reaches 97.29%, which is 1.98% higher than the original YOLOv4mAP value, and the detection of eye parts is increased by 6%. In terms of fatigue judgment, in view of the real-time problem of fatigue detection, this thesis&nbsp;judges by the duration of fatigue action. When the model is detected, the face is first detected, and then the features of the eyes and mouth are detected. If fatigue features are detected, frame processing is performed. If the fatigue features last for 2 seconds, the system will issue a fatigue alarm reminder. In addition, in order to better adapt to hardware devices with different performances, model compression technology is used to compress and accelerate the model before model deployment. By comparing the three methods of model pruning, model quantization and TensorRT model compression, the actual performance of this model is Effect, select the TensorRT model acceleration method, which doubles the model weight size and enables the model to be deployed on devices with smaller memory. After the machine is accelerated, the frame rate is changed from 18 to 40, and the frame rate deployed on the hardware is changed from 3 to 9, which meets the real-time and accuracy of fatigue monitoring.</p>