Semantic Segmentation of High-resolution Point Clouds Representing Urban Contexts

Abstract

Point clouds are becoming an increasingly common digital representation of real-world objects, and they are particularly efficient when dealing with large-scale objects and/or when extremely high-resolution is required. The focus of our work is on the analysis, 3D feature extraction and semantic annotation of point clouds representing urban scenes, coming from various acquisition technologies, e.g., terrestrial (fixed or mobile) or aerial laser scanning or photogrammetry; the task is challenging, due to data dimensionality and noise. In particular, we present a pipeline to segment high-resolution point clouds representing urban environments into geometric primitives; we focus on planes, cylinders and spheres, which are the main features of buildings (walls, roofs, arches, ...) and ground surfaces (streets, pavements, platforms), and identify the unique parameters of each instance. This paper focuses on the semantic segmentation of buildings, but the approach is currently being generalised to manage extended urban areas. Given a dense point cloud representing a specific building, we firstly apply a binary space partitioning method to obtain small enough sub-clouds that can be processed. Then, a combination of the well-known RANSAC algorithm and a recognition method based on the Hough transform (HT) is applied to each sub-cloud to obtain a semantic segmentation into salient elements, like façades, walls and roofs. The parameters of primitive instances are saved as metadata to document the structural element of buildings for further thematic analyses, e.g., energy efficiency. We present a case study on the city of Catania, Italy, where two buildings of historical and artistic value have been digitized at very high resolution. Our approach is able to semantically segment these huge point clouds and it proves robust to uneven sampling density, input noise and outliers.