Research On Fatigue Driving Detection Based On Deep Learning

In recent years,a significant increase in the standard of living per capita has been attributed to the growing power of China overall,which has led to an increasing number of private car ownership.As a result,the occurrence of road safety accidents has become more frequent,with fatigue driving causing a relatively large number of accidents.In order to reduce accidents,more and more researchers are focusing on fatigue detection.However,the traditional fatigue driving detection method based on the characteristic posture of the vehicle during driving is easily affected by factors such as road conditions and the driving habits of the driver,resulting in low accuracy.In addition,although the detection method based on the physiological signal characteristics of the driver is more accurate,its devices are more complex,and such devices need to be worn by the drivers,which may affect the normal driving.It is well known that facial information such as eyes,mouth and head movement trajectory are closely related to the state of drivers.Therefore,we focus on the eyes and head movement trajectory of the drivers based on deep learning fatigue detection method in this thesis to judge the current fatigue state of the drivers.The advantage of this method is that a high accuracy rate can be achieved without requiring the driver to wear the relevant instrumentation.Moreover,it is feasible to detect the fatigue characteristics of both the eyes and head movements of the drivers at the same time,which can effectively reduce the error caused by the occasional nature of a single characteristic.In this thesis,we use Retina Face and Res Net neural networks for face target detection and feature point detection around the eyes,respectively.The novelty of the thesis is to propose the use of a model of rotational head to obtain a single eye of the drivers for detection,and to create a dataset containing 2596 single eye images.Firstly,we proposed a fatigue detection algorithm that weighted fusion of the information of eyes for the drivers and the information entropy corresponding to the head motion trajectory.Then the algorithm proposed in this thesis was verified using 20 videos,and the average accuracy reached 96.84%,which showed that the algorithm could accomplish the detection task.Finally,the testing process of this thesis is summarized,and a client software is created using PyQt5.