Consistent Roof Geometry Encoding for 3D Building Model Retrieval Using Airborne LiDAR Point Clouds

2017 ISPRS International Journal of Geo-Information  
A 3D building model retrieval method using airborne LiDAR point clouds as input queries is introduced. Based on the concept of data reuse, available building models in the Internet that have geometric shapes similar to a user-specified point cloud query are retrieved and reused for the purpose of data extraction and building modeling. To retrieve models efficiently, point cloud queries and building models are consistently and compactly encoded by the proposed method. The encoding focuses on the
more » ... geometries of building roofs, which are the most informative part of a building in airborne LiDAR acquisitions. Spatial histograms of geometric features that describe shapes of building roofs are utilized as shape descriptor, which introduces the properties of shape distinguishability, encoding compactness, rotation invariance, and noise insensitivity. These properties facilitate the feasibility of the proposed approaches for efficient and accurate model retrieval. Analyses on LiDAR data and building model databases and the implementation of web-based retrieval system, which is available at http://pcretrieval.dgl.xyz, demonstrate the feasibility of the proposed method to retrieve polygon models using point clouds. extraction and efficient cyber city modeling. Based on the concepts of data reuse and crowsourcing, the proposed system can efficiently construct a 3D city model which is one of key components in virtual geographic environment. Content-based model retrieval methods can be classified into two categories, model-based retrieval and view-based retrieval, based on the shape descriptor used in encoding. In model-based retrieval, shape similarities are measured using various geometric descriptors, including shape distribution [6], shape spectral [8], and shape topology [9] . In methods based on shape distribution, geometric features defined over feature spaces are accumulated in bins [6] . A histogram of these bins is utilized as the signature of a 3D model. In shape spectral methods, geometric shapes are transformed to a spectral domain and the spectral coefficients are used in shape matching and retrieval [8] . In topology-based methods, model topologies are represented as skeletons, and retrieval is performed based on the assumption that similar shapes have similar skeletons [9] . Unlike model-based methods, view-based methods represent 3D geometric shapes as a set of 2D projections, and 3D models are matched using their visual similarities instead of geometric similarities [7,10,11]. Each projection is described by image descriptors, and shape matching is reduced to measurement of similarities among the views of the query object and models in the database. Model-based and view-based methods perform well on existing benchmarks for polygon model encoding and retrieval. However, these methods are not designed for unorganized, noisy, sparse, and incomplete point clouds. Recently, Chen et al. [12] proposed point cloud encoding using a set of low-frequency spherical harmonic functions (SHs). With the preprocessing of data resampling and filling, the approach can alleviate the difficulties caused by sparse and incomplete sampling of point clouds. However, the use of low-frequency SHs decreases the ability to distinguish objects with similar geometric shapes thereby leading to ambiguity in shape description. To improve shape distinguishability, an roof geometry encoding that integrates shape distribution with visual similarity is proposed. The main idea is to represent point clouds and polygon models using top-view depth images that can describe the shapes of building roofs and avoid disturbances from insufficient sampling of building side-views. The depth images are further encoded by geometry features with spatial histograms, which introduce the properties of compact description, rotation invariance, noise insensitivity, and consistent encoding of point clouds and polygon models. These properties lead to a compact storage size and real-time retrieval response time. Furthermore, the visual similarity in depth images and shape distribution in spatial histograms increase the distinguishability of geometric shapes. The remainder of this paper is organized as follows. Section 2 describes the methodology of point cloud encoding and building model retrieval. Section 3 introduces the properties of the proposed encoding method. Section 4 discusses experimental results, and Section 5 presents conclusions and future work. Methodology System Overview
doi:10.3390/ijgi6090269 fatcat:p5bl7tecdzhyjpno3wt7eapdzu