The Research On Driving Status Of Drivers Based On Multi-information Fusion

In recent years,with the popularity of automobiles,traffic accidents have occurred frequently,especially in recent years,"two type of passengers car and one type of dangerous car" accidents have brought huge loss of life and property to people.According to the survey,more than 80% of traffic accidents are caused by drivers' distraction.In December 2018,the security commission of the state council proposed in the "notice on promoting the application of intelligent video surveillance and alarm technology" that all vehicles with " two type of passengers car and one type of dangerous car " should be equipped with intelligent video surveillance system to realize automatic identification and real-time alarm of drivers' unsafe driving behaviors to ensure safe driving.Therefore,this paper proposes a multi information fusion method based on driver's driving action information and face fatigue state information.The system is divided into three parts.In the first part,the driver's behavior and actions are classified by the key information of the driver's skeleton,including left hand lifting,right hand lifting,normal limbs,normal head state,left head turning,right head turning,bow head,smoking,drinking water and making phone calls.In the second part,the driver's fatigue driving level is detected by the key point information of the driver's face,including normal(T1),moderate(T2)and severe(T3).The third part combines the information of the first two parts to design a driving behavior risk level discriminator to complete the real-time detection of dangerous driving behavior.The main research work of this paper includes the following four aspects:First,for the detection of specific targets in the cab,this paper introduces the development of YOLOv3 network and its advantages in real-time target detection.In order to improve the detection accuracy and speed of three types of objects of specific size in the cab.This paper makes full use of the characteristics of the size of the objects to be detected.It removes the detection layer of large objects in the original yolov3 network,and uses K-means ++ method to re-cluster anchor.Compared with the original YOLOv3 network,the detection accuracy of YOLOv3 Plus network has increased from 91.8% to 98.9%,and the detection speed has increased from 40 fps to 58 fps.Secondly,in the detection of driver's attitude,based on the 12 key points obtained from Open Pose network,this paper proposes a driver's attitude classifier based on SOM(self organizing map)+ HL(Hebbian learning).In this paper,firstly,based on the original SOM network,the learning of data sensitivity curve is added to make the neuron better express the driver's action details,and the original Hebbian learning rules are improved to avoid the situation that only the weight increase is allowed when the neuron is activated.Compared with the original SOM + HL classifier,the accuracy of the improved classifier for 10 types of driver behavior detection is improved from 77.57% to 94.42%.Thirdly,in the detection of driver's fatigue state,this paper first completed the location of 24 key points on the driver's face through the improved Le Net network,and proposed a fuzzy inference system for driver's fatigue level detection,that is,first through the spatial position relationship of key points on the face to determine the threshold value of twinkling and mouth opening,and then combined with PERCLOS value and fatigue driving experience knowledge,blink and yawn frequency information,and finally complete the classification of driver fatigue level.Fourthly,in the design of the driver's dangerous driving behavior classification and discrimination system,this paper uses the method of multi information fusion to comprehensively judge the driver's dangerous driving behavior.The system integrates driver fatigue driving state information,driver driving behavior information and driver experience information,and finally designs a retrieval form of driving behavior risk classification,which completes the classification of driver dangerous driving behavior by the way of look-up table voting.Finally,the detection accuracy of the system for three levels of driver's driving behaviors,namely,low-risk driving(R1),medium-risk driving(R2)and high-risk driving(R3),is 91.62%(R1),92.75%(R2)and 97.56%(R3),which can well meet the detection requirements.