基于概率图模型的机载LiDAR点云数据分类

发布时间：2018-01-16 02:30

本文关键词：基于概率图模型的机载LiDAR点云数据分类　出处：《中国地质大学(北京)》2017年硕士论文　论文类型：学位论文

【摘要】：由于机载LiDAR点云数据具有场景复杂、目标丰富、密度不均匀以及含有噪声等特点,使得现有三维场景目标提取识别方法自动化程度不高、识别准确率不高等特点。针对目前三维场景中目标识别率低的问题,本文在综合信息航空影像和机载点云数据各自特点的基础上,延伸了多源数据融合的思想,提出一种融合航空影像光谱信息辅助点云分类的贝叶斯网络模型,将三维场景分类为地面、低植被、高植被和建筑物四类。随着遥感技术的不断发展,诸多学者致力于将遥感技术应用于电力线监测中,其中机载LiDAR以其独特的优势,也被日渐广泛地应用于电力巡线中。针对目前电力线场景点云分类自动化程度低的问题,文中提出了一种基于随机森林后验概率的马尔科夫随机场模型,用于电力线场景的点云分类。本文在综合分析机载LiDAR点云数据特点及场景的情况下,对城区以及电力线巡线场景的LiDAR点云进行分类,主要研究内容如下:(1)对于城区的LiDAR点云数据,本文在分析点云数据特点和总结前人经验的基础上提取了用于描述点及其邻域形状特征的几何特征。对航空影像进行多尺度分割,并分析影像对象提取相应的光谱信息,将其赋予对应的点,以此辅助点云的分类。基于互信息理论分析提取的点云几何特征和影像对象特征之间的依赖关系,以此为基础构建最优的贝叶斯网络模型,用于描述点云分类特征向量的联合概率分布,实现点云数据的自动分类和信息提取。(2)对于电力线巡线场景的LiDAR点云,文中结合空间金字塔理论构建了多尺度视觉分类特征以此描述空间点及其邻域的几何形状信息;接着利用随机森林分类器描述观测数据的概率分布,然后基于马尔科夫随机场模型建立顾及上下文信息的先验概率,从而构建一个多标记能量函数;最后为了优化分类标签配置,利用多标记图割技术使构建的能量函数最小化。实验结果表明,通过将影像对象信息融合到点云分类过程中,能有效地改善各类地物的分类正确率,分类正确率能达到90%以上;基于MRF模型的优化,通过考虑空间关系来优化分类结果。与其他传统分类器而言,本文提出的贝叶斯网络模型和MRF模型在大多数数据上要优于其他常规分类器。
[Abstract]:Because the airborne LiDAR point cloud data have the characteristics of complex scene, rich target, uneven density and noise, so the existing 3D scene target extraction and recognition methods are not high degree of automation. The recognition accuracy is not high. Aiming at the problem of low target recognition rate in 3D scene, this paper extends the idea of multi-source data fusion on the basis of synthesizing the characteristics of information aerial image and airborne point cloud data. This paper presents a Bayesian network model which integrates spectral information of airborne image to assist point cloud classification. The 3D scene is classified into four categories: ground, low vegetation, high vegetation and buildings. With the development of remote sensing technology, the classification of 3D scene can be divided into four categories: ground, low vegetation, high vegetation and buildings. Many scholars are committed to the application of remote sensing technology in power line monitoring, in which airborne LiDAR has its unique advantages. In order to solve the problem of low automation degree of power line scene point cloud classification, a Markov random field model based on the posteriori probability of stochastic forest is proposed in this paper. Point cloud classification for power line scene. This paper classifies the LiDAR point cloud of urban area and power line patrol scene by synthetically analyzing the characteristics and scene of airborne LiDAR point cloud data. The main research contents are as follows: (1) for the urban LiDAR point cloud data. In this paper, based on the analysis of point cloud data and previous experience, the geometric features used to describe points and their neighboring shape features are extracted, and multi-scale segmentation of aerial images is carried out. The image objects are analyzed to extract the corresponding spectral information, and the corresponding points are assigned to assist the classification of point clouds. Based on the mutual information theory, the dependence between the extracted point cloud geometric features and the image object features is analyzed. On this basis, an optimal Bayesian network model is constructed to describe the joint probability distribution of the feature vectors of point cloud classification. Realize the automatic classification of point cloud data and information extraction. 2) for the power line survey scene LiDAR point cloud. In this paper, the multi-scale visual classification features are constructed based on the spatial pyramid theory to describe the geometric shape information of spatial points and their neighbors. Then the random forest classifier is used to describe the probability distribution of observed data, and then a priori probability with context information is established based on Markov random field model, and a multi-marker energy function is constructed. Finally, in order to optimize the collocation of classification labels, the energy function is minimized by using multi-label graph cutting technique. The experimental results show that the image object information is fused into the point cloud classification process. It can effectively improve the classification accuracy of all kinds of ground objects, and the classification accuracy rate can reach more than 90%. Based on the optimization of MRF model, the classification results are optimized by considering the spatial relationship. The Bayesian network model and the MRF model proposed in this paper are superior to other conventional classifiers in most data.
【学位授予单位】：中国地质大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P23

【相似文献】