PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
April 2016
257
Automatic Point Cloud Registration Using a
Single Octagonal Lamp Pole
Ting On Chan, Derek D. Lichti, David Belton, and Hoang Long Nguyen
Abstract
Registration is an essential procedure for merging point clouds de-
fined in different coordinate systems associated to different scan-
ner positions and orientations. It is usually the first step before the
point clouds are further processed to provide spatial information
of a scene to support engineering applications. In this paper, a
new automatic registration method based on a novel geometric
model of a polygonal object is presented. Since the cross section
of the shaft of many lamp poles is octagonal, registration based on
an octagonal pyramid model is proposed. The presented method
only requires a single, common octagonal lamp pole observed in
both point clouds, though actual overlap of the point clouds is not
strictly required. It can be performed as long as the model param-
eters can be estimated by fitting the point observations to the mod-
el. Moreover, no user interaction is needed to derive approximate
values, so the proposed registration can be completely automated.
Three independent datasets captured by two scanners were used
to verify the method. The registration accuracies in the horizontal
and vertical directions were up to 11.7 mm and 4.4 mm at ap-
proximately 62 m and 17 m away from the scanner, respectively.
With such high accuracies, the estimated registration parameters
can serve as a set of initial parameters for fine registration using
algorithm such as the iterative closest point (
ICP
).
Introduction
Terrestrial laser scanners (
TLS
s) are advanced, three-dimen-
sional (3D) survey instruments that capture the surfaces of
objects as highly redundant sets of 3D point coordinates
known as point clouds. Point cloud registration is an import-
ant procedure for merging groups of point clouds defined in
different Cartesian coordinate systems. The rigid body trans-
formation parameters for transforming multiple point clouds
into a common coordinate system must be estimated.
Registration is widely used in many areas to facilitate the
use of point clouds. For example, a complete point cloud of
an electrical substation is obtained by registering multiple
single scans (Al-Durgham and Habib, 2014) to facilitate the
subsequent spatial analysis of electrical insulators (Ara-
stounia and Lichti, 2013). The point clouds captured by
terrestrial and airborne lidar systems are registered to deliver
a large point cloud rich with both roofs and façades (Wu
et
al
., 2014; Teo and Huang, 2014). A full point cloud for a giant
structure is created by registering approximately one hundred
of scans (Chang
et al
. 2012) for structural analysis.
The point cloud registration often begins by obtaining a set
of initial values for the transformation parameters through a
coarse registration procedure. The accuracy of the initial pa-
rameters is refined using advanced estimation methods such as
the iterative closest point (
ICP
) method (fine registration). The
initial values are often obtained by a manual selection of com-
mon target points, which is time consuming and lacks standard
selection criteria. Some research efforts have been made to
improve this aspect. Instead of using manually-selected points,
methods have been proposed to automatically identify and
match conjugate planes, lines, curves, and other features such
as cylinders from individual static point clouds for registration.
Rabanni
et al
. (2007) proposed two registration methods
(indirect and direct methods) that use various geometric
primitives often found in industrial scenes such as planes,
spheres, cylinders, and tori. For the indirect method, point
clouds of the defined features are extracted automatically.
This is followed by feature parameter estimation with least-
squares model fitting. The registration parameters are then
computed by minimizing the sum of squared differences
between feature parameters of conjugate features from two
scans. For the direct method, the estimated parameter values
from the indirect method serve as initial values for the regis-
tration parameter adjustment. The registration parameters are
estimated in such a way that the sum of the squared orthogo-
nal distance between points and their corresponding geomet-
ric models of the features is minimized.
For urban environments rich with planar features, Brenner
et al
. (2008) proposed two registration methods that estimate
initial parameter values. The first method is a plane-based
approach that first automatically segments the planar features
from the scenes and then uses the normal vectors as primi-
tives to compute the registration parameters in closed form.
The correct matches between conjugate planes are determined
using a voting algorithm that computes the similarity between
different planes using the discrepancy between their normal
vectors and orthogonal distances from the origin. The second
method is an iterative cell alignment scheme inherited from
the field of robotics. The method first converts the master
point clouds as cubic cells. The slave point cloud is divided
into many horizontal slices. Then, the slices are repeatedly
aligned to the cells to compute the similarity scores based on
the point distribution probability within each individual cell.
The registration parameters are determined by checking the
highest score obtained from the iterative alignment process.
Theiler
et al
. (2014) estimate the initial values of the regis-
tration parameters by first extracting some key points from the
point clouds as primitives. The key points were computed by
several statistical methods inherited from conventional image
processing techniques. The points can be treated as heavily
down-sampled version of the original point cloud with most
geometric characteristics preserved. Then, the 4-points con-
gruent sets (4
PCS
) algorithm is used to match conjugate key
Ting On Chan and Derek D. Lichti are with the Department of
Geomatics Engineering, University of Calgary, 2500 Universi-
ty Drive NW, Calgary, Alberta T2N1N4, Canada
(
).
David Belton and Hoang Long Nguyen are with the Depart-
ment of Spatial Sciences, Curtin University, GPO Box U1987,
Perth, WA 6845, Australia.
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 4, April 2016, pp. 257–269.
0099-1112/16/257–269
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.4.257
