3D Building Model Retrieval System Using Airborne LiDAR Point Clouds

Abstract

With the development of Web 2 0 applications and scanning equipment an increasing number of three-dimensional (3D) building models have been made available on web-based model-sharing platforms Based on the concept of data reuse a building model on the Internet is retrieved and reused for modeling instead of reconstructing a new model from point cloud through a complex and nontrivial process namely model-driven or data-driven modeling A 3D building model retrieval system is proposed in this study to realize this data reuse concept The system can retrieve similar building models from a database by using a point cloud acquired through airborne LiDAR The proposed system consists of two main steps namely point cloud classification and model retrieval aims to efficiently retrieve building models that are similar to the input point cloud in terms of shape First this study focuses on building model extraction and accurate classification of LiDAR point clouds which comprise of fundamental and critical step for the separation of different objects In the point cloud classification geometric features that are generally utilized in the separation of different objects play an important role in successful classification Among the geometric features eigen-features calculated through the principal component analysis are the commonly used geometric features; they can describe the local geometric characteristics of a point cloud However eigen-features calculated through the principal component analysis of a covariance matrix are sensitive to LiDAR data with inherent noise and incomplete shape sampling because of the non-robust statistical analysis To obtain reliable eigen-features from LiDAR data and improve classification accuracy this study introduces a method of analyzing the local geometric characteristics of a point cloud through the use of a weighted covariance matrix with a geometric median rather than the standard covariance matrix and the sample mean which are sensitive to point distribution In this method each point in the neighborhood of a point is assigned a weight to represent its spatial contribution in the weighted principal component analysis and to estimate the geometric median which can be regarded as a localized center of a shape A LiDAR point cloud can be accurately classified with a reliable covariance matrix and geometric median and the point clouds belonging to building models can be extracted Second motivated by the concept of data reuse an encoding approach is proposed for 3D building model retrieval through the use of LiDAR point clouds The key to a successful model retrieval system is the accurate and efficient representation of a 3D shape The basic idea behind the proposed method is to represent point clouds and building models with a complete set of spherical harmonics (SHs) SH is a compact and simple shape descriptor that has the advantages of reducing storage size and search time In addition SH representation is insensitive to noises if only the low-frequency SHs are employed The inherent rotation-invariant property of SH encoding enables the retrieval system to address the problem of 3D rotate-transform between the point cloud and building models The multi-resolution nature of SH encoding also allows for the efficient matching and indexing of the model database Furthermore a data filling and re-sampling approach is proposed to solve the problem of incomplete shapes of point clouds and the aliasing problems of SH coefficients attributed to sparse sampling of point clouds In the experiments of point cloud classification qualitative and quantitative analyses on airborne LiDAR data and simulated point clouds show a clear improvement of the proposed method with improved eigen-features compared with that of standard eigen-features The classification accuracy is improved by 1 6% to 4 5% through the use of a supervised classifier In the experiment of model retrieval system qualitative and quantitative analyses of LiDAR data show the clear superiority of the proposed method over other related model retrieval methods

Date of Award	2015 Aug 27
Original language	English
Supervisor	Chao-Hung Lin (Supervisor)