A Supervoxel Segmentation Method Based On Feature Line Extraction Of Point Cloud Data

With the increasing application of point cloud data in the 3D scene,more and more researches are devoted to processing point cloud data in different scenes.Supervoxel segmentation of point clouds is a method of over-segmentation of point cloud data.In most cases,as a preprocessing method of other graphics methods which processes point clouds,it can provide geometric regular and meaningful over-segmentation results for subsequent processing,which can save time and space costs in subsequent steps.In this thesis,a supervoxel segmentation method based on feature line extraction of point cloud data is proposed,which can make the supervoxel segmentation matches the feature lines,and keeps the regular segmentation shape in the nonfeature regions.This method firstly initializes the point cloud data and extracts the feature lines of the point cloud data.Then,seeds are spread in both sides of the feature lines symmetrically to make the boundaries of the supervoxels centered on these seed points coincide with the feature lines and evenly spread the points in the non-feature regions.Finally,under the constraint of the fixed seed points around the feature lines,the restricted Voronoi cells of all the other seed points are generated and their positions are optimized in Lloyd's method iteratively.At the end of the iteration,according to the positions of seed points,the point cloud data is segmented and clustered.In this thesis,through an initialization operation,the feature lines of the original point cloud data are extracted.According to the feature lines,the boundaries of the supervoxels align with the boundaries of the real object,and the segmentation result and the visualization effect are very good in the regions around the features.The segmentation result of this method depends on the feature line extraction.The better the feature lines fit the feature edges and boundaries of the object,the better the segmentation result.In this method,the Lloyd method is used to iteratively optimize the positions of the seed points,and the local surface information of point clouds is considered so that the distribution of seed points has good geometric properties,the shape of supervoxels is nearly regular,compact,and it converges fast.This algorithm is suitable for point cloud data with obvious geometric features,such as outdoor scenes,buildings and so on.In this thesis,a large number of experiments are carried out on the specified data,and the experimental results show that the method has excellent performance in the segmentation on the feature regions,and obtains a good visual effect of the segmentation results.This thesis also compares the proposed method with other existing methods,and discusses the advantages and disadvantages of this method,as well as the work that can be carried out in the future.