Simplification Of Locomotive Running Gear Beased On Three Dimensional Point Cloud

The condition monitoring and fault diagnosis of locomotive running part is the key point of routine maintenance of locomotive depot,and it is a powerful guarantee to reduce the serious events caused by locomotive running part fault.The image detection of locomotive running part is still dominated by manual visual inspection,which is affected by subjective factors such as the working state,working experience,psychological quality of maintenance personnel,as well as objective factors such as light,oil pollution,shooting angle,etc.With the development of 3D scanning technology,it is of great practical significance to study the locomotive on-line detection system based on the surface defect and 3D shape measurement technology.However,the 3D model needs a lot of data to express the original data format,which not only increases the difficulty of surface reconstruction,but also reduces the subsequent processing such as 3D model reconstruction.The goal of this thesis is to extract reliable key points from the scattered point cloud with large data and noise information to represent the information of the whole 3D object model,and to simplify the point cloud data of locomotive running part through the distribution of key points.1.Since the pointcloud data of locomotive bearing cover obtained by three-dimensional laser scanner is scattered,messy and huge,and there is no obvious geometric topological relationship between points.Combined with the characteristics of the target point cloud data,this thesis uses k-d tree to establish the geometric relationship of the measured data,and uses hierarchical clustering to cluster the point cloud globally in the space domain,so as to establish an appropriate point cloud model Combined geometric and topological relations,improve the efficiency of subsequent simplification.2.In order to solve the problem that the traditional key point detection algorithm is sensitive to noise and depends on the shape characteristics of the object model,this thesis proposes a multi-scale key point extraction algorithm based on normal weighting.By introducing the normal weighted shape index value of local surface to describe the change information of adjacent surface and measure the difference of three-dimensional points,the noise data is smoothed to a certain extent,which makes the algorithm less affected by noise.Experimental results show that compared with other key point extraction algorithms,this method reduces the sensitivity of the algorithm to noise,improves the efficiency of human-computer interaction,and improves the detection ability of key points in the plane area.3.The point cloud data collected by the laser 3D scanning equipment has high density and large amount of data,which contains too much redundant information,and the effect of simplification directly affects the quality of subsequent surface reconstruction.In view of the randomness of clustering center selection of traditional K-means based point cloud reduction algorithm,which leads to the instability of clustering results,this thesis proposes a hierarchical clustering based adaptive K-means 3D point cloud reduction.The algorithm uses the sum of squares of deviation to measure the similarity of each cluster,sets the maximum number of hierarchical clusters as the clustering cut-off condition to establish a bottom-up hierarchical clustering tree,and uses the number of non empty leaf nodes and clustering centroid as the basis of adaptive K-means clustering of three-dimensional point cloud.The traditional K-means method is improved in the selection of the initial clustering centroid.And K-means clustering is subdivided again.After the subdivided point cloud is obtained,it is mapped to Gauss sphere for mean shift clustering.Finally,the center of gravity is used to represent the point cloud of clustering cluster,and then the simplified point cloud is obtained.Compared with the traditional K-means algorithm and the random sampling method,the sample distribution after the output of this method depends on the density distribution of the feature points.Therefore,for areas with smoother features,fewer points are reserved,and for areas with more complex features,more points are reserved,which greatly improves the accuracy of the algorithm.