Quintuple Local Coordinate Images for Local
Shape Description
Wuyong Tao, Xianghong Hua, Ruisheng Wang, and Dong Xu
Abstract
Owing to poor descriptiveness, weak robustness, and high
computation complexity of local shape descriptors (
LSDs
),
point-cloud registration in the case of partial overlap and ob-
ject recognition in a cluttered environment are still challeng-
ing tasks. For this purpose, an
LSD
is developed in this article
by proposing a new local reference frame (
LRF
) method and
designing a novel feature representation. In the
LRF
method,
two weighting methods are applied to obtain robustness to
noise, point-density variation, and incomplete shape. Ad-
ditionally, a vector representation is calculated to disambigu-
ate the sign of the x-axis. The feature representation encodes
the local information by generating the local coordinate
images from five views. Thus, more geometric and spatial
information is included in the descriptor. Finally, the perfor-
mance of the
LRF
method and the
LSD
is evaluated on several
popular data sets. The experimental results demonstrate
well that the
LRF
is robust to noise, point-density variation,
and incomplete shape, and the
LSD
holds strong robustness,
superior descriptiveness, and high computational efficiency.
Introduction
In computer vision and computer graphics, local shape
description has been extensively researched and proven to
be very successful. A wide variety of applications are imple-
mented by means of local shape description, such as registra-
tion of point clouds (Guo
et al.
2014; Dong
et al.
2018; Yang,
Xiao and Cao 2019), object recognition (Johnson and Hebert
1999; Zhong 2009; Aldoma
et al.
2012), and simultaneous
localization and mapping (Tong and Barfoot 2013; Zhang and
Singh 2014). With the advent of laser scanning devices, 3D
point-cloud data became increasing popular. The attention
of scholars has been moved from 2D descriptors to 3D ones
(Knopp
et al.
2010; Darom and Keller 2012). In comparison
with 2D descriptors, 3D descriptors are not influenced by
illumination condition or shadow, by which more accurate
rigid transformation (i.e., translation and rotation) could be
obtained. Nevertheless, 3D descriptors still need to be further
studied for point-cloud registration in the case of partial over-
lap and object recognition in cases with occlusion and clutter.
The existing 3D descriptors can be broadly classified into
two categories: global and local. A limitation of global shape
descriptors is that the object needs to be separated from a
cluttered scene beforehand, because they encode the geomet-
ric information of the entire object. Therefore, they cannot be
used for incomplete objects, due to their sensitivity to occlu-
sion and clutter. By contrast, local shape descriptors (
LSDs
)
encode the geometric information of the local neighborhood
around the key point into a feature vector representation, so
they are robust to clutter and occlusion. Consequently,
LSDs
are suitable for point-cloud registration with partial overlap
and object recognition in chaotic scenes.
LSDs
are computed on the key points extracted from point
clouds. Through comparison of the similarity of the
LSDs
,
the key points in the model point cloud are then paired with
those in the scene point cloud. If the descriptors of two key
points from different point clouds are similar (the same in
principle), the correspondence between the two key points
is established. Therefore,
LSDs
should be invariant to rigid
transformation. However, owing to the effect of different kinds
of nuisances (including noise, outliers, varying point density,
occlusion, clutter, and missing regions), the calculated
LSDs
are hard-pressed to keep the invariance. This requires that the
LSDs
have the ability to resist these nuisances (i.e., robustness).
Another important characteristic is descriptiveness. A good
LSD
should describe as much information of the local surface
as possible to obtain high descriptiveness. This is helpful to
improve descriptor matching performance, which directly in-
fluences the accuracy of estimated transformation parameters
and the computational time in verification and refinement
phases. A large number of
LSDs
have been devised for differ-
ent purposes. Examples contain snapshots (Malassiotis and
Strintzis 2007), spin images (SI; Johnson and Hebert 1998),
fast point feature histograms (
FPFH
; Rusu, Blodow and Beetz
2009), rotational projection statistics (
RoPS
; Guo
et al.
2013),
triple orthogonal local depth images (
TOLDI
; Yang
et al.
2017c),
and signature of histograms of orientations (
SHOT
; Tombari,
Salti and Di Stefano 2010b). These descriptors can be gener-
ally divided into two classifications, in light of whether a
local reference frame (
LRF
) method is employed or not. The
descriptors without
LRF
—such as SI,
FPFH
, and local surface
patch (Chen and Bhanu 2007)—use only the local geometric
information to form feature vectors. This kind of descriptor
takes geometric information into account but discards spatial
information; hence, they have limited descriptiveness (Tom-
bari
et al.
2010a). In contrast, the descriptors with
LRF
(e.g.,
RoPS
,
SHOT
, and Tri-Spin-Image; Guo
et al.
2015) have higher
descriptiveness because both spatial and geometric informa-
tion are encoded into feature vectors with respect to the
LRF
. A
comparative analysis of
LSDs
can be found in Guo
et al.
(2016).
Wuyong Tao is with the School of Geodesy and Geomatics,
Wuhan University, Wuhan 430079, China; and the
Department of Geomatics Engineering, University of Calgary,
Calgary T2N 1N4, Canada.
Xianghong Hua is with the School of Geodesy and Geomatics,
Wuhan University, Wuhan 430079, China (
.
edu.cn).
Ruisheng Wang is with the School of Geographical Sciences,
Guangzhou University, Guangzhou 510006, China; and the
Department of Geomatics Engineering, University of Calgary,
Calgary T2N 1N4, Canada (
).
Dong Xu is with the State Key Laboratory of Information
Engineering in Surveying, Mapping and Remote Sensing,
Wuhan University, Wuhan 430079, China.
Photogrammetric Engineering & Remote Sensing
Vol. 86, No. 2, February 2020, pp. 121–132.
0099-1112/20/121–132
© 2020 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.86.2.121
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2020
121
