Research And Application Development Of Intelligent Precision Measurement Algorithm Based On 3D Point Cloud Registration Technology

Under the heavy load and the influence of the external environment,the steel rail can produce various defects and wear.At the same time,long train lines and complex terrain,which cause low manual detection efficiency and seriously affect the safety of driving.However,the intelligent precision measurement based on machine vision can not only achieve digital management,but also can provide efficient work guarantee for rail maintenance,and it can significantly keep the railcars’ safety and make the railcars stablely operating.To sum up,this dissertation studies the 3D measurement algorithm based on point cloud registration technology and the surface defect detection model,and realizes the related algorithm and software engineering design.The specific work and research results of the dissertation are as follows:(1)Research on the preprocessing algorithm of rail point cloud data.The amount of point cloud data acquired through a three-dimensional scanning instrument is large,and there is no certain topological relationship between the points.And during the scanning process,the noise points will be generated due to the influence of object vibration.This dissertation establishes the topological relationship between these points based on KD-Tree,and uses conditional filtering and bilateral filtering to filter out the noise of point cloud data.At the same time,the algorithm tries the best to simplify the date without changing the data characteristics of the point cloud.(2)Research and implementation of 3D point cloud registration algorithm.In order to solve the problem that the feature of rail surface is not obvious,this dissertation proposes a point cloud registration algorithm based on auxiliary calibration ball。The algorithm extracts the center coordinates of the auxiliary calibration ball and solves the rotation parameters of the point clouds from different perspectives to realize the rail point cloud data registration.At the same time,aiming at the problem that the registration accuracy of super4pcs algorithm is not high on the whole,and the overlap ratio of two point clouds has to be large enough,a point cloud registration algorithm based on super4pcs is proposed.The point cloud is divided by clustering and segmentation by feature and distance information.The corresponding relationship is determined according to the overall characteristics between classes to extract the overlapping area of two point clouds.Finally,the multi-view point cloud registration is completed through the overlapping area.Simulation experiments prove that compared with the traditional FPFH algorithm、SCA-IA algorithm and Super4PCS algorithm,the algorithm proposed in this dissertation effectively improves the accuracy of point cloud registration.(3)Designing and implementint the intelligent precision measurement system for rail.On the basis of the conventional detection parameters,the system calculates and analyses the cross-sectional profile shape parameters of the 60kg/m rail.At the same time,the profile parameters such as the side radius of the rail head and the 45° angle wear of the rail head that cannot be measured by traditional measurement methods are solved,and the profile is also realized.The shape parameters are visualized,the relevant measurement curves are drawn,the corresponding measurement reports are generated,and the measurement errors of the system are quantitatively analyzed.Finally,it is proved that the results of this dissertation can meet the technical requirements of measurement accuracy by experiments.(4)Research and implementation of rail defect detection model based on deep learning.Aiming at the over fitting problem caused by the small number of negative samples in rail sample data,a dynamic data augmentation method is proposed to simulate the real scene.Combined with the morphological characteristics of rail surface defects,a surface defect detection model based on fast r-cnn is proposed.The algorithm proposed in this dissertation has good performance in the task of surface defect recognition,especially for micro defects such as scars and spots.The simulation results show that the detection accuracy of this algorithm is better than that of faster r-cnn model in many categories.The detection accuracy of scar is 91.8%,and mAP is 81.2%.