Research On Intelligent Detection,Inspectionand Positioning Of Foreign-Body On High-speed Pavement Based On Binocular Stereo Vision

Pavement Foreign-Body is one of the main contents in the daily work of the expressway inspection and maintenance department.The high-speed driving environment is relatively closed,with high speed and large flow.The occurrence location of Foreign-Body has a certain randomness,which is easy to pose a threat to road traffic safety.Generally,the safety threat caused by the Foreign-Body in the overtaking lane is higher than that in other lane positions.Therefore,various objects in different lanes should have different exclusion response degrees.It is of great significance to find the existence of Foreign-Body in time during patrol inspection and take elimination measures to ensure driving safety and maintain high-speed safety environment.At present,manual way is widely used in the inspection of spills.However,with the growth of highspeed total mileage,the manual method will gradually be unable to meet the needs of inspection in the future.The existing research is trying to use automatic detection methods to realize the rapid detection of ForeignBody,but it cannot really be combined with the actual inspection requirements,and cannot provide substantive reference and guidance for the current inspection work.Combined with the requirements of Foreign-Body detection,this paper proposes a set of intelligent detection and mileage positioning method of pavement spillage based on binocular stereo vision and yoov5 network model.This method can realize the automatic identification of all kinds of Foreign-Body in expressway environment,the two-dimensional grid positioning of horizontal pavement and the automatic measurement of the size of Foreign-Body.The experimental data comes from multisource heterogeneous data collected in the expressway environment.The data includes foreground image,depth image,global positioning system(GPS)and inertial measurement unit(IMU)sensor information.The research contents include: 1)Based on the collected road foreground images,the training set of Foreign-Body is constructed,the weight parameters of the network structure are initialized by the method of transfer learning,and the training strategy of neural network model is improved to train the Foreign-Body detector model suitable for highspeed scene;2)A mileage location method combining binocular stereo vision and GPS is proposed to locate the mileage of objects thrown along the road;3)A clustering method of lane line is proposed to determine the lane position of the Foreign-Body,and finally realize the two-dimensional grid positioning of the spilled object on the horizontal pavement;4)Combined with the application scenario of high-speed inspection operation,an intelligent inspection system for Foreign-Body on highspeed pavement is developed for automatic detection of Foreign-Body.Through the inspection test on the 2098 km to 2079 km section of G15 Shenhai Expressway in Fuqing City,Fuzhou,the detection accuracy and recall rate of Foreign-Body under all sizes are 0.82 and 0.92 respectively.The mileage positioning error is within 6m and the lane positioning accuracy is about 0.934.It has certain application value for the detection and inspection positioning of Foreign-Body in the actual scene.