CLASSIFICATION OF LIDAR DATA OVER BUILDING ROOFS USING K-MEANS AND PRINCIPAL COMPONENT ANALYSIS

Abstract

The classification is an important step in the extraction of geometric primitives from LiDAR data. Normally, it is applied for the identification of points sampled on geometric primitives of interest. In the literature there are several studies that have explored the use of eigenvalues to classify LiDAR points into different classes or structures, such as corner, edge, and plane. However, in some works the classes are defined considering an ideal geometry, which can be affected by the inadequate sampling and/or by the presence of noise when using real data. To overcome this limitation, in this paper is proposed the use of metrics based on eigenvalues and the k-means method to carry out the classification. So, the concept of principal component analysis is used to obtain the eigenvalues and the derived metrics, while the k-means is applied to cluster the roof points in two classes: edge and non-edge. To evaluate the proposed method four test areas with different levels of complexity were selected. From the qualitative and quantitative analyses, it could be concluded that the proposed classification procedure gave satisfactory results, resulting in completeness and correctness above 92% for the non-edge class, and between 61% to 98% for the edge class.