A Real-time Fatigue Detection System Based On Embedded Devices

Fatigued driving often leads to traffic accidents and ranks among the top causes of traffic accidents.Generally speaking,fatigue driving detection can be divided into three steps: face detection,eye detection,and fatigue recognition.After the camera inputs the captured image,the driver’s face first is extracted from the image in the face detection stage,then the eye detection is accurately extracted from the face in the eye detection stage,and finally,in the fatigue recognition stage whether the driver is fatigued is judged from the extracted eyes.Face detection is a kind of target detections.Faster R-CNN,as the most representative detection method,can significantly improve target detection precision.However,as the accuracy increases,the computational complexity also increases.These models that can only be trained on GPU are difficult to be used in the industrial practice.As we all know,the operation efficiency of CPU is only 1 / 10 of that of GPU.So CPU is a common choice in the industry.In this context,SSD,and YOLO algorithms are proposed to improve the running speed of deep target detection.However,through experiments,it is found that the running speed of these methods on the Raspberry pie is only 2-3 fps,in which the gap between CPU and GPU and the performance gap between raspberry pie and computer CPU are the reasons for the slow running speed.The speed of the depth target detection model determines whether it can be applied in the industrial field.In this context,aiming at the particularity of cab photos,this paper improves the SSD algorithm and enhances the running speed to 10.06 fps only with the 0.78% accuracy loss.Considering that the face in the driver image is large and only a single face needs to be recognized,the improvements of this paper include the following three parts:1.Introduce the lightweight network MobileNet＿V2 to replace the VGG in the original SSD algorithm as the backbone;2.Abandon the small target detection feature map with large time-consuming in target detection;3.Cancel the time-consuming non-maximum suppression(NMS)algorithm,and directly take the feature window with the largest probability as the detection result.The face detection part has recognized a complete face.On this basis,it is a relatively simple task to detect the position of human eyes.In this paper,machine learning algorithm is used to save running time and meet real-time requirements.Finally,the eye detection part uses the Gentle Adaboost algorithm,which has a higher accuracy than the traditional classification algorithm by introducing the concept of bagging.In addition,this paper uses the Camshift algorithm for human eye tracking,which can effectively reduce the trigger frequency of human eye detection algorithm,and improve the real-time performance of the overall system.After detecting the position of the human eye,this paper uses the Percentage of Eyelid Closure Over the Public Over Time(PERCLOS)method to recognize the fatigue driving on the recognized face.After multiple rounds of experiments,PERCLOS method P80 standard is selected as the human eye closure standard,and the threshold values of awake,mild fatigue,fatigue and severe fatigue are established with the P80 score and the blink frequency.Finally,through the establishment of cameras,raspberry pie,batteries and loudspeaker part,an anti fatigue driving system is built.