Modeling Indoor Spaces Using Decomposition
and Reconstruction of Structural Elements
Ruisheng Wang, Lei Xie, and Dong Chen
Abstract
We present a framework for modeling architectural structure
of indoor scene given unorganized 3D point clouds collected
from an indoor mobile laser scanning system. The proposed
pipeline employs a decomposition-and-reconstruction
strategy to create the indoor structural elements on the floor
plan. In contrast with existing methods, our method can
deal with both inner and outer wall reconstruction through
solving the optimization energy problem using graph cuts. In
addition, our approach does not rely on timestamp informa-
tion encoding line-of-sight information from each point to
the laser scanner. It allows our algorithm to process most of
the data sets. In particular, the proposed alpha-shape based
door reconstruction and a hierarchical clustering technique
is able to identify each room without prior information
regarding the total number of rooms. Our experiments on
real and synthetic point cloud scenes show that our ap-
proach is robust and accurate for modeling indoor spaces.
Introduction
Recent years have witnessed an increased demand for recon-
struction and modeling of indoor scenes (Choi
et al
., 2015)
due to its potential applications in building information
modeling (
BIM
), indoor mapping and navigation, architecture,
and more. Current practice in 3D indoor model generation
is either a manual or interactive tedious process. Automated
reconstruction of indoor scenes at all levels of the modeling
stages is desired. Compared to the outdoor reconstruction, the
reconstruction of architectural elements of building interior
such as walls and doors is still in an early stage, but draws
more and more attention from photogrammetry, computer vi-
sion and computer graphics communities.
Indoor point cloud can be collected by stationary terrestri-
al laser scanning (
TLS
). The
TLS
scanner is normally equipped
onto a tripod to scan a portion of the indoor scene on each
station. Therefore, a large scene can be covered by a set of
scanning from different viewpoints. However, the
TLS
data ac-
quisition method suffers from low mapping efficiency because
of the laborious station resetting procedure. The continuous
scanning system for indoor mobile mapping is called mobile
laser scanning (
MLS
) which integrates laser scanner, initial
measurement unit (
IMU
) and optical camera together; and the
equipment is fixed onto a mobile platform (e.g., trolley). The
MLS
system can continuously get point clouds for the indoor
scenes along the data acquisition route which is gradually be-
ing widely used in the indoor environment scanning.
In this paper, we focus on utilizing point clouds from
MLS
to reconstruct the indoor building models. The assumption
of our modeling approach is fairly general: all the walls are
perpendicular to the floor and ceiling, but the intersecting
angles between walls can be arbitrary. Given unorganized
point clouds and the corresponding trajectory of the trolley,
our method can automatically generate a 2D floor plan, detect
different doors, identify each individual room, and create 3D
building models. Given that our method is based on some pre-
vious works, we explicitly specify our uniqueness as follows:
1. We propose an automatic pipeline for reconstruction of
building modeling from indoor
MLS
point cloud through
a decomposition-and-reconstruction strategy. Our method
does not need timestamp information to determine light-
of-sight between each point and the laser scanner, which
makes the proposed approach better adapted to most of
data sets.
2. Our method can enhance the model representations
through structural component reconstruction: wall, door,
and room. This method can deal with both inner and outer
wall reconstruction using a binary labeling process and
minimize the reformulated energy function through the
graph cut optimization. In particular, the proposed two
algorithms regarding door and room elements reconstruc-
tion significantly improve the detail of building models.
Related Work
From the data source acquisition point of view, the indoor
modeling methods can generally be categorized into recon-
struction methods from indoor
MLS
points and reconstruction
methods from stationary
TLS
points.
Indoor MLS Reconstruction
Sanchez and Zakhor (2012) introduce an approach that di-
vides indoor points into ceilings, floors, walls, and other small
architectural structures according to the normal computed
from the Principle Component Analysis (
PCA
) algorithm.
Then, the Random Sample Consensus (
RANSAC
) plane fitting
is employed to detect planar primitives. Nevertheless, they
utilize fragmentized planar primitives to express the building
rather than a set of mesh faces. Xiao and Furukawa (2012) gen-
erate 2D Constructive Solid Geometry (
CSG
) models then stack
over these models through a greedy algorithm to create 3D
CSG
models. During the processing of the 2D
CSG
models, they
Ruisheng Wang is with Beijing Advanced Innovation Center
for Imaging Technology, Capital Normal University Beijing,
China, 100000; and the Department of Geomatics Engineering,
University of Calgary,2500 University Drive NW, Calgary,
Alberta, Canada, T2N 1N4 (
).
Lei Xie is with the Department of Geomatics Engineering,
University of Calgary, 2500 University Drive NW, Calgary,
Alberta, Canada, T2N 1N4 (
).
Dong Chen is with College of Civil Engineering, Nanjing
Forestry University,159 Longpan Road, Nanjing, Jiangsu,
China, 210037; and the Department of Geomatics Engineering,
University of Calgary, 2500 University Drive NW, Calgary,
Alberta, Canada, T2N 1N4.
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 12, December 2017, pp. 827–841.
0099-1112/17/827–841
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.12.827
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2017
827
