IEEE Geoscience and Remote Sensing Magazine - December 2020 - 40

Multispectral lidar performs well in extracting water, vegetation,
and shadows, but the data are not easily obtained [32], [33].
TLS, also called static LS, scans with a tripod-mounted stationary sensor. Since it is used in a middle- or close-range environment, the point cloud density is very high. Its advantage
is its ability to provide real, high-quality 3D models. TLS has
been commonly used to model small urban or forest sites and
document heritage sites or works of art. Semantic3D.net [34]
is a typical TLS benchmark data set.
MLS systems operate from moving vehicles, usually cars.
A current hot topic for research and development is autonomous driving, for which HD maps are essential. The generation of HD maps is therefore the most significant application
for MLS systems. Several mainstream point cloud benchmark
data sets are captured with MLS systems [35], [36].
ULS systems are usually deployed on drones or other unmanned vehicles. Since they are relatively cheap and very flexible, this recent addition to the lidar family is gaining popularity.
Compared to ALS systems, where the platform is also above the
objects, ULS systems can conduct lidar surveys from a shorter
distance. They can thus collect denser point clouds with higher
accuracy. Because the platform is compact and lightweight, ULS
systems offer high operational flexibility. Therefore, in addition
to traditional lidar tasks (e.g., acquiring DSMs), ULS systems offer advantages for conducting agriculture, forestry, and mining
surveys and for monitoring disasters [37]-[39].
Since the system is always moving with the platform, it is
necessary for LS to combine the positions of points with GNSS
and inertial measurement unit data to ensure a high-quality
matching point cloud. Lidar has been the most important data
source for point cloud research and has been used to compare
and evaluate the quality of point clouds from other sources.
RGB-D POINT CLOUD
An RGB-D camera can acquire both RGB and depth information. There are three kinds of RGB-D sensors, each based on a
different principle: 1) structured light [40], 2) stereo [41], and 3)
time of flight [42]. Similar to lidar, the RGB-D camera measures
the distance between the camera and the objects. But the camera
generates pixel-wise depth data, rather than unstructured points.
An RGB-D sensor is much cheaper than a lidar system. Microsoft's Kinect is a well-known and widely used RGB-D sensor
[40], [42]. In an RGB-D camera, relative orientation elements
between or among different sensors are calibrated and known,
so coregistered, synchronized RGB images and depth maps can
be easily acquired. Obviously, the point cloud is not the direct
product of RGB-D scanning. But, since the position of the camera's center point is known, the 3D space position of each pixel
in a depth map can be easily obtained and then directly used
to generate the point cloud. RGB-D cameras have three main
applications: object tracking, human pose or signature recognition, and SLAM-based environment reconstruction. Since
mainstream RGB-D sensors are for close-range applications,
much closer even than those for TLS systems, they are usually
employed in indoor environments. Several mainstream indoor
PCSS benchmarks use RGB-D data [43], [44].
40

SAR POINT CLOUD
Interferometric SAR (InSAR), a radar technique crucial for remote sensing, generates maps that show surface deformations
or digital elevations based on comparisons of multiple SAR image pairs. InSAR-based point clouds have demonstrated their
value over the past few years and are opening up new possibilities for point cloud applications [45]-[49]. SAR tomography
(TomoSAR) and persistent scatterer interferometry (PSI) are
two major techniques that generate point clouds with InSAR,
extending the principle of SAR into the 3D realm [50], [51].
TomoSAR's advantage over PSI is its ability to enable detailed
reconstruction and monitoring of urban areas, especially of human-made infrastructures [51]. The TomoSAR point cloud has
a point density comparable to that of ALS lidar [52], [53]. Especially useful for applications in building reconstruction in urban areas, these point clouds have the following features [46]:
1) TomoSAR point clouds reconstructed from spaceborne
data have a moderate 3D positioning accuracy on the order
of 1 m [54], enabling decimeter-level accuracy by geocoding error-correction techniques [55]. By comparison, ALS
lidar provides accuracy typically on the order of 0.1 m [56].
2) Due to their coherent imaging nature and side-looking geometry, TomoSAR point clouds emphasize different objects
with respect to lidar systems. The side-looking SAR geometry enables TomoSAR point clouds to possess rich façade
information ([57] presents results using pixelwise TomoSAR for the high-resolution reconstruction of a building
complex with a very high level of detail from spaceborne
SAR data). Temporarily incoherent objects, e.g., trees, cannot be reconstructed from multipass spaceborne SAR image
stacks. To obtain the full structure of individual buildings
from space, façade reconstruction using TomoSAR point
clouds from multiple viewing angles is required [45], [58].
3) Complementary to lidar and optical sensors, SAR is so far
the only sensor capable of providing fourth-dimension
information from space, i.e., temporal deformation of the
building complex [59], and microwave scattering properties of the façade reflect geometrical and material features.
InSAR point clouds have two main shortcomings that affect their accuracy: 1) due to limited orbit spread and the small
number of images, the location error of TomoSAR points is
highly anisotropic, with an elevation error typically one or two
orders of magnitude higher than in range and azimuth; and 2)
due to multiple scattering, ghost scatterers may be generated,
appearing as outliers far away from a realistic 3D position [60].
Compared with the aforementioned image-derived, lidarbased, and RGB-D-based point cloud, data from SAR systems
have not yet been widely used for studies and applications.
However, mature SAR satellites, such as TerraSAR-X, have collected abundant global SAR data, which are available for InSARbased reconstruction at a global scale [61]. Hence, SAR point
clouds can be expected to play a conspicuous role in the future.
POINT CLOUD CHARACTERS
As sensors were developed and various applications
emerged, point clouds evolved in three stages: 1) sparse
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

DECEMBER 2020

http://www.Semantic3D.net

IEEE Geoscience and Remote Sensing Magazine - December 2020

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2020