Recognition Of Key Components From Train Bottom Based On 3D Point Cloud

The recognition of train bottom key components is a necessary guarantee for the safety of train in operation.As for the limitation of acquisition environments and the shooting angles,the recognition methods based on two-dimensional images cannot work well sometimes.In contrast,3D point cloud data would not be affected by the scale change or the environmental light,besides,point clouds could provide more geometric features than 2D images.So it can greatly improve the recognition of train bottom key component by using the point clouds.Because the train components point clouds have a huge number of points,an efficient point cloud down-sampling algorithm should be used for reducing the number of points with a high speed,while keeping the features as much as possible.So the research on efficient down-sampling algorithms is one of the key parts in this thesis.Establishing an effective local feature descriptor and using an accurate key point matching algorithm are another two crucial tasks in recognition based on the3 D point cloud.Since the descriptors need to keep enough descriptive ability against the effect of noise,occlusion,and incomplete regions in the point cloud,a suitable key point matching algorithm can get more precise matched pairs,which can improve the accuracy of the surface matching,identification and positioning.In order to make the recognition performance better,the following work are conducted in this thesis:1.Two commonly used sampling algorithms,Random Sampling(RS)and Iterative Farthest Points Sampling(FPS),are evaluated,each of which has a suitable applicated condition.According to the characteristics and requirements in sampling of train component point clouds,an efficient point sampling algorithm is proposed,combining with the octree and point density based on FPS.Experimental results show that the running time is greatly reduced,and the calculation efficiency is significantly improved.2.The methodology of Critical Point Layer(CPL)sampling algorithm is introduced,using the deep learning network.Considering that the maxpooling only retains the points that contribute to the largest feature,a new critical point layer sampling algorithm is proposed based on CPL.This sampling algorithm can retain more important points that make contributions to the features in each dimension,retain the top three points in the eigenvalue ranking of each dimension in features,and output the sampling points according to the contributions in features.Combined with this new critical point layer with point cloud classification network,the experiments on the Model Net40 dataset show that the overall accuracy and average accuracy of the classification have been improved.3.To build an effective descriptor,a Multi-Statistics Histogram Descriptor is proposed,combining the spatial distribution and geometric attributive features.This method has been evaluated based on Stanford 3D dataset.The experimental results demonstrate the superiority of Multi-Statistics Histogram Descriptor because its descriptive ability and robustness to noise and mesh resolution are greater than that of carefully selected baselines(e.g.,FPFH,SHOT,Ro PS,and Spin Image descriptors).4.In the stage of key point matching,in order to reduce the number of wrong matched corresponding key points,a new key point matching algorithm is developed,which could identify more corresponding point pairs.It has been verified that the error of rotation and translation matrix is much smaller based on the proposed key point matching algorithm,and the precise corresponding point pairs can be captured,resulting in the enhanced recognition and registration for three-dimensional surface matching.