Research And Implementation Of Fatigue Driving Detection Based On Deep Learning

With the gradual improvement of people’s living standards,comfort and convenience have become the primary condition for people to choose food,clothing,housing and transportation.In the aspect of "transportation",private cars have gradually become a means of transportation for the public to travel,and the traffic flow in festivals such as National Day has doubled,and the occurrence of traffic accidents is also on the rise.According to the analysis of the reasons,the accidents caused by drivers’ fatigue and driving after drinking accounted for the highest proportion.Among them,drunk driving behavior can be controlled by alcohol content detector,while fatigue driving is more controlled by drivers themselves.Therefore,the design of a fatigue driving detection system is particularly important in real life.Based on the completion of the research project,this thesis independently developed a set of non-contact,real-time and adaptive fatigue driving detection prototype system by comparing and referring to relevant literatures.First adopt the method of gradient direction histogram(HOG)of face detection,in order to reduce the actual driving the complex interference environment,through the methods of gray level is changed as the image preprocessing before detection,in the face detection experiments to test the images under different illumination conditions,the results showed that the method of HOG still maintain good accuracy in terms of illumination change.Then this thesis proposed an improved algorithm based on PFLD(FLHPD algorithm)to facial landmark detection and head pose estimation,retained the original algorithm,the advantages of simple and rapid at the same time,considering the need to use the head pose estimation and PFLD fatigue detection algorithm in such aspects as close their eyes and looked up and a drop in accuracy problem,The corresponding improvements are made,including redesigning the network structure,modifying the loss function and solving the problems of sample imbalance.Usually,drivers are in a state of fatigue,in the face and head show the corresponding fatigue characteristics,such as closing time increases,yawning,sleepy nodded,to detect the fatigue characteristics,must first eyes,mouth,head,which can identify different states of the it is important to the selection of threshold,aiming at this problem,put forward a set of reasonable threshold selection scheme.And it has been verified by experiments.With the method of multi-index fusion criterion after fatigue,real-time computing unit time duration PERCLOS,blinking frequency and close my eyes,yawning frequency and sleepy nodding frequency fatigue parameter values.Among them,the thresholds of blink frequency and head attitude Angle are usually calculated according to the driving video of the driver in the awake state,which makes the fatigue detection system self-adaptive.In this thesis,Python language is used to realize the specific detection algorithm,and Wx Python is combined to realize the UI interface of the system.Finally,the driving video files are tested by the system.The experimental results show that the fatigue detection and early warning system designed in this thesis can meet the actual needs.