Contents I. Introduction
With the advancement of computer vision technology and the emergence of lidar, consumer-grade RGB-D cameras, and binocular cameras, 3D point cloud technology has been widely used [1]. Point cloud segmentation is to cluster points into a uniform area according to the local features of the point cloud, and apply it to various aspects of scene analysis according to the segmentation results. The general method is to construct a graph based on the grid of the input point cloud, and use information such as the normal, smoothness or bumpiness of the boundary line to perform cluster segmentation. Segmentation is a key link in point cloud data processing, which can provide important information for subsequent surface reconstruction and feature extraction, and is the premise and basis for object recognition. The commonly used point cloud segmentation methods include: region growing [2], [3], concave-convex segmentation [4], watershed algorithm [5], hierarchical clustering [6], spectral clustering [7], etc. The color region growing algorithm was first proposed by Zhan and Xiao [8]. The region segmentation algorithm was first applied to two-dimensional image processing. The segmentation steps are as follows: first, find a seed pixel for each region to be segmented as the starting point of growth, and then use the same or similar properties as the seed pixel in the neighborhood around the seed pixel. Pixels (determined according to some pre-determined growth or similarity criterion) are merged into the region where the seed pixel is located. Continue the above process with these new pixels as new seed pixels, until there are no more pixels that meet the conditions to be included. The region growth segmentation algorithm in the point cloud uses the normal or the curvature as the judgment condition, and sets the curvature difference threshold and the normal angle threshold, and the growth area is formed when the conditions are met. The traditional region segmentation algorithm is simple and easy to implement, but the noise in the data and the uneven distribution of geometric features will have a great impact on the effect of the algorithm. Chen J et al.[9] propose a learnable region growing method for class-independent point cloud segmentation. The proposed method is able to segment objects of any class using a single deep neural network without any need for any preprocessing in their shape and size. However, the method takes a long time and requires high hardware. Zhang W et al.[10] proposed a new algorithm for region growing based on 2-D-3-D mutual projection. The initial seed points are selected by the geometric information of the point cloud in 3-D space and the projection plane, and then the visibility of each point is estimated according to the defined growth criterion. The results show improvement in accuracy, but the scope of application is limited. Xin Chen et al.[11] used the multi-level spline surface fitting method to optimize the region growing algorithm to realize the segmentation of lung lobe point cloud data, but this method requires a large amount of computing memory and takes a long time. By estimating the curvature of point cloud data, Li et al.[12] set the minimum curvature point as the seed point, which improved the accuracy and reliability of point cloud segmentation, but their method could not achieve good results for areas with small curvature changes.
