IEEE Robotics & Automation Magazine - December 2015 - 120

of the acquisition viewpoints are available (in which case
the initial estimate only needs to be refined) as opposed to
the situation where they are unknown (in which case an
initial alignment needs to be computed solely using the acquired data).
In addition to these differences, which often require
adopting ad hoc solutions, one can also exploit certain characteristics of the data to
improve speed or robustness. Since a fundamental
Publicly available software
problem of registration is
to find the most suitable
libraries and data sets are
algorithm with an appropriate parameterization,
a way to promote such
we dicuss three different
examples ranging from
research and to make
dense RGB-D image sequences for object modelresults comparable.
ing to high-resolution and
low-resolution laser range
scans for the outdoor mapping of buildings and cities.
Large-Scale 3-D Laser Scans
Commercially available 3-D laser scanners, such as those
used by surveying technicians, acquire high-resolution
large-scale scans that usually consist of several million
points. With a comparably ample maximum measurement
range, these sensors can be used to capture large scenes
such as buildings and streets. To create a complete model
(e.g., a whole city or an urban district), the acquired scans
need to be aligned and aggregated. Frequently, surveying
technicians use additional means, such as special reflectors, that assist in obtaining a fine alignment of the data.
In our example, we assume that such information is not
available and that we want to 1) compute an initial alignment for a pair of large-scale 3-D laser scans, and 2) refine
Listing 10. Multiscale keypoint and descriptor
computation.
#include 
#include 
(...)
FPFHEstimation::Ptr fest (new (...));
fest->setInputCloud (cloud)
fest->setInputNormals (normals);
MultiscaleFeaturePersistence fper;
std::vector keypoints;
std::vector scale_values = { 0.5f,
l.Of, 1.5f };
fper.setScalesVector (scale_values);
fper.setAlpha (1.3f);
fper.setFeatureEstimator (fpfh_estimation);
fper.setDistanceMetric (pci::CS);
fper.determinePersistentFeatures (*features,
keypoints);

120

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

DECEMBER 2015

the initial estimate to obtain a perfect overlay of the acquired data.
Initial Alignment
As mentioned in the "Coarse Alignment via Descriptor
Matching" section, the PCL provides implementations for a
variety of 3-D descriptors that can be used for initial alignment. Here, the key issue is not only to find an appropriate
descriptor for the problem at hand, but also an appropriate
scale, i.e., the support size on which to compute each descriptor. Of particular interest in this context is the distinctiveness
of each descriptor that influences the number of possible ambiguities in the matching stage as well as the repeatability,
such that the descriptor computed on the same scene point
over different scans looks the same. A particularly robust approach is to compute the same class of descriptor at various
scales and to focus the registration on points and associated
descriptors that prove to be distinctive at the same scale, and
are persistent over multiple scales. The PCL provides joint estimation components that finds such keypoints and descriptors (" MultiScaleFeaturePersistence).
Listing 10 shows an example for computing keypoints
for an input point cloud and the respective persistent feature descriptors. First, the selected descriptor (in our example we use FPFH [24]) is fetched to an instance of the

Listing 11. Initial alignment using keypoints
and descriptors.
#include 
#include 
#include  (...)
Correspondences::Ptr correspondences (new
Correspondences);
CorrespondenceEstimation cest;
cest.setInputSource (source_features);
cest.setInputTarget (target_features);
cest.determineCorrespondences
(*correspondences);
Correspondences::Ptr corr_filtered (new
Correspondences);
CorrespondenceRejectorSampleConsensus
 rejector;
rejector.setInputSource (source_keypoints);
rejector.setInputTarget (target_keypoints);
rejector.setInlierThreshold (2.5);
rejector.setMaximumIterations (1000000);
rejector.setRefineModel (false);
rejector.setInputCorrespondences
(correspondences);;
rejector.getCorrespondences (*corr_
filtered);
TransformationEstimationSVD trans_est;
trans_est.estimateRigidTransformation
(*source_keypoints, *target_keypoints,
*corr_filtered, transform);
Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2015
