ABSTRACT Individual roof plane extraction from Light Detection and Ranging (LiDAR) point-cloud data is a complex and difficult task because of unknown semantic characteristics and inharmonious behaviour of input data. Most of the existing state-of-the-art methods fail to detect small true roof planes with exact boundaries due to outliers, occlusions, complex building structures, and other inconsistent nature of LiDAR data. In this paper, we have presented an improved building detection and roof plane extraction method, which is less sensitive to the outliers and unlikely to generate spurious planes. For this, a robust outlier detection algorithm has been proposed in this paper along with a robust plane-fitting algorithm based on M-estimator SAmple Consensus (MSAC) for detecting individual roof planes. Using two benchmark datasets (Australian and International Society for Photogrammetry and Remote Sensing benchmark) with different numbers of buildings and sizes, trees and point densities, we have evaluated the proposed method. Experimental results show that the method removes outliers and vegetation almost accurately and offers a high success rate in terms of completeness and correctness (between 80% and 100% per-object) for both roof plane extraction and building detection. In most of the cases, the proposed method shows above 90% correctness.

中文翻译：

---

从 LiDAR 数据中提取建筑物屋顶的异常值检测和鲁棒平面拟合

摘要 由于未知的语义特征和输入数据的不协调行为，从光检测和测距 (LiDAR) 点云数据中提取个体屋顶平面是一项复杂而艰巨的任务。由于异常值、遮挡、复杂的建筑结构和 LiDAR 数据的其他不一致性质，大多数现有的最先进方法无法检测到具有精确边界的小型真实屋顶平面。在本文中，我们提出了一种改进的建筑物检测和屋顶平面提取方法，该方法对异常值不太敏感，并且不太可能产生虚假平面。为此，本文提出了一种鲁棒的异常值检测算法以及基于 M 估计器 SAmple 共识 (MSAC) 的鲁棒平面拟合算法，用于检测单个屋顶平面。使用具有不同建筑物数量和大小、树木和点密度的两个基准数据集（澳大利亚和国际摄影测量与遥感协会基准），我们评估了所提出的方法。实验结果表明，该方法几乎准确地去除了异常值和植被，并且在屋顶平面提取和建筑物检测的完整性和正确性（每个对象在 80% 到 100% 之间）方面提供了很高的成功率。在大多数情况下，所提出的方法显示出 90% 以上的正确率。实验结果表明，该方法几乎准确地去除了异常值和植被，并且在屋顶平面提取和建筑物检测的完整性和正确性（每个对象 80% 到 100% 之间）方面提供了很高的成功率。在大多数情况下，所提出的方法显示出 90% 以上的正确率。实验结果表明，该方法几乎准确地去除了异常值和植被，并且在屋顶平面提取和建筑物检测的完整性和正确性（每个对象在 80% 到 100% 之间）方面提供了很高的成功率。在大多数情况下，所提出的方法显示出 90% 以上的正确率。