联合分布约束的激光雷达点云空间上下文建模与分类

Classification is the basic and critical step for global change, resource survey and other researches of geoscience. This study focusing on LiDAR point cloud data, designs a novel classification method which takes contextual information into consideration. Using the confidence interval of statistics, a more stable point cloud entity representation method objected to point cloud classification is first proposed. The method extracts three kinds of entities, i.e., smooth surface, rough surface and individual point. Then, the contextual relationships are studied and refined from the point and entity levels. In the procedure of refinement, the analysis of variance (ANOVA) is employed to exclude the useless and misleading context. At last, joint distribution is utilized to model the contextual information, and is embedded into the training and prediction procedures in the tree-based model to implements the contextual classification of LiDAR point cloud data. Three open-accessed LiDAR datasets are employed to perform the experimental analysis, compare with the existing methods, and demonstrate the advantages of our proposed algorithms.

数据分类是全球变化、资源调查等地学研究的基础和首要步骤。本课题拟面向激光雷达点云数据，设计一种顾及空间上下文信息的分类方法。首先拟利用统计量的置信区间，设计一种更加稳健的面向点云分类的点云实体表达算法，提取三种实体结构，即光滑表面、粗糙表面和单个点。进而，从点和实体层次，深入挖掘和精化空间上下文关系；其中精化过程拟采用方差分析原理，剔除冗余的、误导性的上下文信息。最后，拟利用联合分布为上下文信息建模，并将联合分布嵌入到树结构分类模型的训练和预测过程中，实现顾及空间上下文的点云分类。本课题拟采用三组开放获取的激光雷达点云数据对以上算法进行实验分析，并与现有方法进行对比，证明本课题所提出算法的优越性。

本项目面向激光雷达点云数据，设计了一种顾及空间上下文信息的点云分类方法。该方法首先充分分析了三维点邻域局部空间的几何属性，采用稳健PCA算法和局部分配策略，对激光雷达点云进行多实体表达。然后，将实体空间联通关系以联合分布的形式进行建模，该联合分布被嵌入到树结构分类模型的训练和预测过程中，实现点云场景的初始分类。最后，利用方差分析原理，深度建模实体上下文关系，在初始分类结果作为节点初始状态的情况下，构建高阶马尔可夫网络，进而通过最大后验状态推理实现点云分类。本项目采用国际摄影测量与遥感学会三维语义标号数据集进行验证，实验证明了本项目方法的有效性。本项目执行情况良好，发表学术论文7篇，培养研究生3名、本科生2名，完成了立项预期目标。