| Visual measurement is a hot research topic in the field of computer vision,with the advantages of being non-contact and highly precise.Compared with 2D image measure-ment,3D point cloud can provide more information about the spatial attitude of the mea-sured target,which ensures higher measurement accuracy and is increasingly favoured by researchers.However,direct measurement of the entire target through point cloud can be impaired by interference from non-critical areas,thus reducing the measurement accuracy.In addition,the huge scale of point cloud also causes considerable resource consumption and low processing efficiency.Therefore,it is a common preprocessing method to use point cloud segmentation to extract the region of interest before perform-ing subsequent calculation,thereby effectively reducing the time and space complexity of the measurement.In this regard,this paper studies point cloud segmentation as a method for extracting key-area point cloud data and measuring wheel-rail contact attitude param-eters.The main contents are as follows:（1）A Transformer-based point cloud segmentation algorithm was proposed.Based on the point cloud neighbor density and neighbor normal vector similarity,three kinds of position encoding were proposed and embedded in the Transformer module as the method of calculating attention.Moreover,hard prediction points at different scales were extracted,attention scores were calculated,and more accurate edge segmentation was achieved through feature fusion.Qualitative and quantitative experimental results showed that the edge segmentation of the proposed algorithm was better than that of the base model in the Shape Net and the wheel-rail dataset,and the m Io U in the wheel-rail dataset was improved by 3 percentage points compared with the base model.（2）A labeled data set of wheel-rail contact attitude was constructed,and an ex-traction algorithm based on the a priori key structures of wheel-rail components was pro-posed.Considering the lack of labeled datasets for wheel-rail point cloud segmentation in the supervised deep learning method,this paper constructed initial point cloud data based on the standard wheel-rail CAD model,and generated corresponding wheel-rail compo-nents by using the characteristics of the wheel-rail rigid body and the cutting algorithm,constructing a labeled dataset of wheel-rail contact attitude.Furthermore,a key linear structure extraction algorithm for shaft components and a key plane structure extraction algorithm for rail and wheel rim structures were proposed.Qualitative experimental re-sults showed that the aforementioned extraction algorithm based on key structures of wheel-rail components can accurately extract key structures such as the straight line of axial,the inner plane of the wheel flange and the plane of the rail base（3）A measurement method for wheel-rail contact attitude parameters based on key structures was proposed.The fine-grained wheel-rail components were extracted through the point cloud segmentation algorithm,and the wheelset roll angle was calculated by using the inner side of the wheel rim and the bottom of the rail.Using the center point of the axis’s oriented bounding box and the center point of the track,the wheelset lateral displacement and wheelset ups and downs were calculated.In addition,the swing angle was calculated based on the key straight line of the extracted axial projection point cloud and the direction of the train.The experimental results on wheel-rail contact attitude data sets of different degrees of sparseness showed that the measurement errors of wheelset roll angle and swing angle were 10-2and 10-3degrees,respectively,and the errors of wheelset lateral displacement and wheelset ups and downs were 10-1mm and 10-2mm,respectively,which met the measurement accuracy requirements. |
PDF Full Text Request