Construction Of Car Body Point Cloud Modeland Surface Defect Detection Based On Binocular Stereo Vision

The rapid development of China’s urban rail transit and the increasing number of rail transit vehicles make vehicle safety operation particular important,and the vehicle maintenance work also increases.In order to reduce the work intensity of vehicle maintenance personnel and improve the quality and efficiency of maintenance,the intelligent inspection technology of rail transportation has become a hot issue for research at present.With the popularity of three-dimensional measurement technology and the limitations of twodimensional measurement,three-dimensional vision-based inspection methods exist in various fields of development to varying degrees.Therefore,this thesis focuses on the construction of 3D point cloud model of urban rail transit vehicle body and 3D visual inspection of vehicle body surface,and the main work is as follows:Firstly,according to the basic layout of the intelligent panoramic inspection system of urban rail vehicles,the 3D vehicle body inspection experiment platform was built,and the point cloud data of the vehicle body was measured based on binocular stereo vision,while the point cloud data pre-processing method was studied,and the original point cloud data was preprocessed based on voxel grid filtering and statistical filtering to achieve a better preprocessing effect,so as to lay the foundation for the construction of the 3D point cloud model of the vehicle body and the detection of surface defects at a later stage.Secondly,the vehicle 3D point cloud model construction method based on multi camera was focused,and the vehicle 3D point cloud model construction method was divided into two parts: single camera inter-frame point cloud stitching and multi-camera point cloud stitching.To address the problems of low applicability and poor results of the existing inter-frame point cloud stitching methods in practical projects,a local feature-based inter-frame point cloud stitching method was proposed,the sampling consistency initial registration algorithm based on FPFH feature was used for rough registration,at the same time,the selection method of corresponding points in the registration process was improved to promote the registration speed.The rough registration results were fitted and the wrong registration results were removed,and the fitted results were used as input to the Tr ICP algorithm for fine registration to achieve inter-frame point cloud stitching.In the multi-camera point cloud stitching part,the multi-camera calibration method was studied,and all the cameras in the vehicle body 3D inspection experimental platform were calibrated,and the external parameters of each camera were obtained,and the coordinate system of each camera was fused under the same coordinate system by rigid body transformation to realize multi-camera point cloud stitching,so as to obtain the vehicle body 3D point cloud model.Thirdly,the 3D point cloud defect detection method was studied.For the two kinds of body surface defects,such as foreign body suspension on the body surface and missing parts at the bottom of the train,the defect detection method based on point cloud registration was adopted to detect them.The redundant and noisy points in the 3D point cloud model were analyzed and removed by voxel grid filtering and statistical filtering,respectively,and smoothed by moving least squares,and the surface reconstruction of the standard point cloud model was performed by greedy projection triangulation algorithm,the registration of the point cloud to be detected with the standard point cloud was performed by ICP algorithm,the deviation analysis of the aligned point cloud model was performed and the identified defect areas were located to ensure the surface reconstruction effect of the 3D point cloud model.Finally,the point cloud model construction method and the surface defect detection method of the vehicle body were verified by the point cloud data collected in the actual site.The causes of registration failure in the process of rough registration were analyzed and the results were corrected,and the effect of this stitching method was verified by comparing the effect of various commonly used stitching algorithms.