【作者】 陈飞；

【导师】 李颖；

【作者基本信息】 吉林大学 ， 计算机科学与技术， 2023， 硕士

【摘要】 点云是一种常用的三维信息表现形式,其由多个孤立的点组成,可以完整和直接地表示三维场景中蕴含的信息。然而,原始的点云数据通常是不均匀和庞大的,且存在一定噪声,这些都给数据处理带来了很大挑战。点云分割技术将具有相似特征的点预先聚成一组显露出了相邻物体的边界,为孤立的点提供了拓扑信息,此外还产生了具有更高层次的几何特征信息,这些额外的信息为后续的表面重建和分类等流程带来了便利。点云分割的传统无监督技术中,基于区域生长的方法是最为广泛使用的技术之一,其原理简单且需要相对较少的参数,能够快速地分割出多个物体。但是,此类方法仍然存在一些缺陷,限制了其分割精度和效率。区域生长类方法普遍以点、体素或超体素作为分割单位,其中基于超体素的方法由于对点云进行了不规则的预分割具有更高的效率和更强的鲁棒性。但是现存超体素分割方法也存在难以对齐边界,受点云密度变化影响等问题。除开分割单位的选择问题,此类方法也缺少适应性强的合并准则,无法应对一些复杂情况。此外区域生长算法是一种贪婪策略,其只考虑局部的连接性,一次错误的合并或分离都可能会导致连续的错误。为了解决上述问题,本文在现有算法的基础上进行融合改进,提出了一种使用多尺度超体素的基于边界限制的点云分割方法。该算法首先对点云进行过分割得到多尺度且和真实物体的边界保持一致的超体素。其次,对邻接的超体素进行相似性度量得到连接权值,作为在后续合并流程中的可连接性判断依据。最后,每个超体素初始化为分割块,根据边缘匹配情况来逐步合并满足条件的邻接分割块直到结束。本文的主要贡献如下。(1)本文提出了一种改进的多尺度超体素分割方法,以平坦度来引导超体素生成过程特征代表点的选择改善了超体素的边界保持能力,同时随表面类型自适应的调整超体素的尺度,提高了特征估计时的抗噪能力,解决原始方法中高密度区域超体素尺度过小无法准确提取特征的问题。(2)本文结合平滑性、切平面距离和平坦度近似约束提出了一种分层次度量超体素连接性的方法,对不同类型连接采取不同的计算标准,使区域的合并更具合理性,减少过分割错误。(3)本文使用一种非连续的方式来生长区域,不需要选择种子点,以边界超体素的匹配情况作为相邻分割块的可融合性判断依据,一定程度上减少了欠分割错误。此外,将边界超体素的配对问题转化为二部图的最大匹配个数问题,引入匈牙利算法来获得全局最优解,提高了算法的鲁棒性。本文分别在模拟、室内和室外数据集下对本文所提出算法进行了测试验证,并与多种代表性算法进行了对比论证。实验表明,本文的算法取得了更为优异的分割精度表现。更多还原

【Abstract】 Point cloud,which consists of numerous discrete points,is a commonly used 3d description method.However,the raw point cloud is usually inhomogeneous,massive and permuted by noise,which induces great challenges for data processing.Therefore point cloud segmentation which divides points with similar features into one group reveals boundaries of neighboring objects,provides topological information of discrete points and produces new top-level geometric features.The derived information facilitates the surface modeling and interpretation in the subsequent processes.Among the unsupervised point cloud segmentation techniques,the region-growing based one requires fewer parameters and is capable of segmenting multiple planes in a short time,which makes it widely popular in 3d point cloud segmentation.However,this method still suffers from some disadvantages that limit its accuracy and efficiency for segmentation.The point,voxel and supervoxel are normally employed as the segmentation units for region growing based methods,among which the supervoxel based method is more efficient and robust,because it oversegment the point cloud in an irregular way previously.However it also suffers from some problems such as difficulty in aligning the boundary with objects that have diverse point cloud density.Apart from the selection of segmentation units,the region growing based method also lacks adaptive and universal merge criteria to deal with some complex situations.Additionally,the region growing algorithm is a greedy strategy,it only considers local connectivity.A wrong combination or separation may lead to continuous errors.In order to solve the above problems,this paper proposes a boundary constraints based point cloud segmentation method that employs the multi-scale supervoxel as segmentation units.The algorithm first oversegment the point cloud to obtain multi-scale supervoxels,which are consistent with the boundaries of real objects.Secondly,we assess the similarity of adjacent supervoxels to get the weight value of connections,which can be utilized as the basis for connectivity judgment in the subsequent merging process.Finally,each supervoxel is initialized as a segment,and the adjacent segments that meet the condition are merged gradually according to the number of paired boundary supervoxels until the end.The main contributions of this paper are as follows.(1)In this paper,an improved multi-scale supervoxel segmentation method is proposed.Flatness is used to guide the selection of feature representative points in the process of supervoxel generation,which improves the boundary adherence of supervoxels.Meanwhile,the scales of supervoxels are adjusted adaptively according to the planarity of surface,which increase the robustness of feature estimation and solves the problem that the scale of supervoxel located in high density area is too small to extract accurate features.(2)This paper proposes a hierarchical method to measure supervoxel connectivity based on smoothness,tangent distance as well as curvature difference.Different calculation standards are adopted for different types of connections,which makes the process of region merging more reasonable and reduces over-segmentation errors.(3)An insequential scheme is employed to grow the regions without selecting the seed points.The number of paired boundary supervoxels is the basis for judging the fusibility of adjacent segments,which reduces the under-segmentation errors.In addition,the matching problem of boundary supervoxels is formulated as the problem of maximum matching for bipartite graphs,and the Hungarian algorithm is introduced to obtain the global optimal solution,which improves the robustness of the algorithm.The algorithms proposed in this paper are tested and verified under simulation,indoor and outdoor datasets,and compared with several representative algorithms.Experimental results show that the algorithm in this paper achieves better segmentation accuracy.更多还原