Point Clouds Local Feature Extraction In Inspection Of Locomotive Key Components

The rapid development of China’s railway transportation industry has increased the demand for the safety of key components of railway locomotives,and the three-dimensional surface inspection technology combined with three-dimensional measurement and computer vision has become a research hotspot in industry.As the key steps of three-dimensional target recognition,key point detection and local feature extraction algorithms determine the quality of the recognition results.An effective description of the target is essential for efficient recognition.Strong description ability,strong resolution ability and strong noise resistance ability are always valuable while choosing a descriptor.Using point cloud to describe 3D model makes the rich description of the details of the object surface,but the large data volume and the high redundancy.Therefore,whether the efficient descriptor can be established becomes the key to improve the recognition rate.Not only that,in the task such as multi-frame point cloud registration,streamlining of feature retention,scene semantic analysis and segmentation are inseparable from local feature description and extraction.In the research of feature extraction and description,research in the thesis focuses on the following contents:1.Computer vision algorithms rely on data,and the preprocessing of point cloud data is the basis for complex feature description.Firstly,establish a KD tree makes the efficiency of neighborhood points query higher.Secondly,establish a covariance matrix for the neighborhood,and determine the coordinate axis direction of the local reference frame(LRF)according to the eigenvectors of the neighborhood covariance matrix.That makes the local features independent of the coordinate axis.Thirdly calculating the normal vector and curvature in the LRF is mostly used in many classic algorithms.At last,the least squares fitting of quadric surfaces on LRF is also a basis for the study of the local characteristics of point clouds.2.The principles and implementation process of ISS,KPQ,Ho NO,LSP,3D-Harris and 3DSIFT key point detection algorithm based on curvature graph is introduced in this thesiswith evaluation on them.Repetition rate,descriptiveness,and time complexity are used as the criteria while designing the experiments,and the conclusion is that the 3D-SIFT algorithm has the best effect under the comprehensive consideration of multiple factors.3.The principles and implementation process of descriptors PFH,FPFH,3DSC,USC,Spin Image,Tri SI,LSP,ISS,TRIFT,SHOT,Ro PS,VD-LSD,and a descriptor combined of spatial distribution characteristics and geometric properties proposed by the thesis are comprehensively tested and evaluated on the two indicators description ability and robustness.and the conclusion comes out that the descriptor proposed in this thesis performs stably on all the 4 datasets chosen in the experiment and has both strong description ability and robustness.4.Key point detection algorithms and local feature descriptors are applicated in this thesis into railway non-destructive testing,which effectively improves the accuracy of point cloud registration,improves the efficiency of modeling while retaining detailed features,effectively completes semantic segmentation of the bottom point cloud of locomotive,and successfully recognized the components in the point cloud such as wheels,bolts.