Research On Dangerous Driving Behavior Recognition Based On Fusion Of Localization And Vision Technology

Self-driving technology development under the background of artificial intelligence and intelligent transportation construction by leaps and bounds,has cloth and other auxiliary driving system at all levels in the family car of upgrading,dangerous driving behavior of the active identification as the core content of the auxiliary driving system,covering all the whole scene of dangerous driving behavior accurate identification technology is still need to break through the difficulties.Therefore,the technical innovation and breakthrough of vehicle-mounted active dangerous driving behavior recognition method under a variety of complex road conditions have great research significance.This paper proposes a dangerous driving behavior recognition method based on the fusion of positioning and vision technology,and collects a large amount of data to verify the feasibility and reliability of the model.First of all,aiming at the problems of low accuracy of existing satellite positioning equipment and unable to meet the requirements of lane-level positioning,RTK differential positioning technology is used to achieve centimeter-level positioning effect,and video shooting module is added in secondary development to achieve the acquisition of required multi-source data.Secondly,the characteristics of speeding,illegal lane occupying and frequent lane changing behaviors were analyzed,and a training model with a lane detection accuracy of94.07% was obtained by comparing and improving the Yolo-v5 s target detection network.Trajectory data and image processing data were used.As the input value of the data fusion algorithm combining Naive Bayes with time-series features and D-S evidence theory,the detection rate of three types of dangerous driving behaviors: speeding,illegal occupation,and frequent lane changes is 95.7%.Then,the deficiencies of the existing detection methods of fatigue driving are summarized.Based on the positioning and vision technology,SLDF(Slope of longitudinal Displacement Fluctuation,Longitudinal displacement fluctuation slope)and facial fatigue features(eye EAR,mouth MAR and head pitch)obtained based on a variety of face correlation recognition algorithms(face detection MTCNN,identity recognition Face Net,face key point detection PFLD)were used as data fusion input values.The SVM data fusion algorithm optimized by quantum particle swarm optimization was applied to obtain the fatigue driving recognition model with 86.8% recognition accuracy.Finally,four drivers and three sections were selected to conduct real vehicle experiments for complex scenes of interference positioning,stable operation of video equipment and normal road conditions,and available driving data for 200 minutes were collected.Respectively according to positioning equipment work unstable mountain,tunnel and trees shade the scenarios and video equipment affected by the strong light,the night,head posture migration scenarios such as speeding,illegal encroachments,frequent lane changes,fatigue driving four kinds of dangerous driving behavior comprehensive verification and analysis of test results,the system has an accuracy of 91% in identifying dangerous driving behaviors,and it meets the real-time requirements of vehicle-mounted detection.