IEEE Robotics & Automation Magazine - December 2015 - 111

processing steps, such as segmentation and object reconstruction, can be applied.
Point cloud registration is a recurring task for a number
of applications related to computer vision, computer graphics, robotic perception, photogrammetry, cultural heritage
modeling, digital archaeology, and architecture. The research
over the past years has been aimed at obtaining accurate
alignments for challenging 3-D data, such as isotropic or
nondistinctive objects, and at making the registration process
as automatic as possible.
Publicly available software libraries and data sets are a
way to promote such research and to make results comparable. This article presents the open source PCL and the tools
available for point cloud registration. PCL is a stand-alone,
large-scale, open project for 3-D point cloud processing. It is
released under the BSD license and contains numerous stateof-the-art algorithms for various applications, including filtering, feature estimation, surface reconstruction, model
fitting, segmentation, and visualization, as well as higher level
tools and applications for performing mapping and object
recognition.
In this article, we focus on the pairwise registration of
3-D point clouds, i.e., aligning two point clouds by estimating the transformation between the two view poses under
which the point clouds have been acquired. Techniques and
tools for tackling sequences of point clouds in a globally coherent fashion (i.e., multiview registration) will be covered
in future publications.
Pairwise registration is usually carried out by means of one
of the several variants of the ICP algorithm. Due to the nonconvexity of the optimization, ICP-based approaches require
initialization with a rough initial transformation to increase
the chance of ending up with a successful alignment. Good
initialization also speeds up their convergence. This article will
cover all stages required for the point cloud alignment-from
coarse alignment based on descriptor matching to the ICPbased refinement-illustrating the tools currently available in
the PCL to carry out each task. Furthermore, we present three
pipelines implemented in the PCL together with typical results
for different types of input data, ranging from small-scale object modeling to outdoor 3-D mapping.
General/Basic Registration Pipelines
Problem Formulation
A point cloud is a data structure P used to represent a collection of multidimensional points p ! R n . In a 3-D point
cloud, the elements usually represent the X, Y, and Z geometric coordinates of an underlying sampled surface. When
more information is available, e.g., color information or information about a local surface normal n or curvature l, the
points p ! P are represented by a longer vector.
Given a source point cloud P with points p ! P and a target point cloud Q with points q ! Q, the problem of registration relies on finding correspondences between P and Q and
estimating a transformation T that, when applied to P, aligns

(a)

(b)
Figure 1. Examples of 3-D registration in different scales and
applications. (a) A registered pair of high-resolution 3-D laser
scans (green and blue) in a city. (Data recorded by Dorit
Borrmann and Jan Elseberg, Robotic 3-D Scan Repository:
http://kos.informatik.uni-osnabrueck.de/3Dscans.) (b) A
registered RGB-D image sequence of toy cars on a table.

all pairs of corresponding points ^ p i e P, q j e Qh . One fundamental problem of registration is that these correspondences
are usually not known and need to be determined by the registration algorithm. Given correct correspondences, there are
different ways of computing the optimal transformation with
respect to the used error metric.
To support processing various types of point clouds in the
PCL, all components are templated. The remainder of this article will use code snippets in the style of Listing 1.

Listing 1. Example code listing and terminology.
// specifying types of input and output data
Algorithm alg;
// providing input data
alg.setInputCloud (input_cloud);
alg.setInputNormals (input_normals);
// computing output
PointCloud output_features;
alg.compute(&output_features);

DECEMBER 2015

IEEE ROBOTICS & AUTOMATION MAGAZINE

111

http://kos.informatik.uni-osnabrueck.de/3Dscans

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2015