Fine Grained Point Cloud Segmentation Based On Octree Voxels

Point cloud segmentation is an important research direction of point cloud.At present,it plays an important role in many fields.How to improve the accuracy of point cloud segmentation and reduce the time of the algorithm is the main research of point cloud segmentation algorithm.The traditional point cloud segmentation algorithm is affected by the noise points seriously resulting in low accuracy,and the traditional segmentation algorithm takes points as the basic unit,resulting in low efficiency.For this reason,this paper proposes a fine-grained point cloud segmentation algorithm based on octree voxels,which can not only avoid the interference of noise points and outliers,but also improve the segmentation accuracy and efficiency of the algorithm.The main work is as follows:Firstly,this paper studied the spatial structure of point cloud,and used a point cloud voxelization method based on variant coding octree.Voxelization of point cloud can greatly speed up the segmentation time of the algorithm and avoid the interference of noise points and outliers.According to the input point cloud,the minimum cuboid bounding box is determined as the root node of the octree,and then the octet recursion is carried out according to the spatial trigram limit as the child node.Taking the depth as the termination recursion condition,the leaf node of the octree is finally obtained as the voxel.The feature of the point cloud contained in the voxel is calculated,and the voxel is divided into primary voxel and secondary voxel according to the number of point cloud in the voxel.At the same time,in the process of voxelization,each node in the octree is encoded independently according to its spatial position to facilitate the subsequent neighborhood search.In this paper,the traditional cube voxel is changed to cuboid voxel,which can better eliminate the influence of uneven data density of point cloud.Secondly,the segmentation process of primary voxels is studied,and proposed an improved region growing segmentation algorithm to split the primary voxels.A region growth segmentation algorithm with the principal element as the smallest unit is implemented.The segmentation algorithm of region growing is realized with the primary voxel as the minimum unit.The residual value of the point cloud data fitting plane in the subject element is used as the seed point selection condition in the process of region growing,so that the spatial smoothness can be considered,so as to start to grow in the smooth plane and reduce the number of iterations of the segmentation algorithm.Considering three spatial constraints:spatial nearest,spatial smoothness and spatial connectivity,and then using Gaussian Kernel function to fuse the features of voxel calculation as the judgment condition of region growing algorithm,the problem of under segmentation and over segmentation can be avoided.In the region growing stage,26 neighbors are used for growth,which can realize the segmentation integrity of different types of data and reduce the segmentation time of the algorithm.If the number of voxels in a growth region is less than the set threshold,all voxels in the region are also classified as secondary voxels.Finally,a secondary voxel refine method is presented to solve the over-segmentation problem.Because the number of point clouds in voxels is too small to accurately calculate voxel features,the refinement process of secondary voxels is added to resist the interference of noise and outliers.Traverse the undivided secondary voxel queue,and perform 18 neighbors search on each voxel traversed to obtain the neighboring queue.If there are segmented voxels in the neighboring queue,the secondary voxels are divided into the same region.If there are multiple segmented voxels in the nearest neighboring queue and belong to different regions,Then,the region to which the secondary voxels are divided is determined by calculating the Euclidean Distance of the center of gravity of the voxels.Repeat this step until the number of undivided secondary voxel queues does not change.At this time,all voxels in the undivided secondary voxel queues are regarded as noise voxels and are not divided.In this paper,a fine-grained point cloud segmentation based on octree voxels is proposed,which realizes the effectiveness and accuracy of point cloud data segmentation through the primary voxel region growing and secondary voxel refine process.By comparing the data of TLS and MLS,the experimental results show that the proposed algorithm can improve segmentation accuracy and resist interference in noise and outliers.The analysis of time complexity shows that the efficiency of the proposed algorithm is also improved compared with the traditional segmentation algorithm.