(XIX-B3, 2012
ning point cloud.
ample concensus: a
1s to image analysis
ions of the ACM,
tprint Simplification
quares Adjustment.
ic Conference
)8. Urban GIS from
ational Archives of
Spatial Information
tomatic 3D building
mmetry and Remote
nation of building
al Data Research,
extraction of dense
lings ASPRS 2011
E. Markelin L,
| of buildings and
s. Remote Sensin,g
of building ground
al Archives of
«XXVII. Part B2,
nethod for building
jution lidar data.
Remote Sensing and
3A, pp. 295-301.
undary tracing and
point clouds.
te Sensing, 73(7),
Development of
exes for building
mote Sens., 43, pp.
automatic building
tion models. /SPRS
1 using planar faces
tional Archives of
patial Information
) building model
| ground plans.
emote Sensing and
pp. 37-43.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
ALTERNATIVE METHODOLOGIES FOR THE ESTIMATION OF LOCAL POINT
DENSITY INDEX: MOVING TOWARDS ADAPTIVE LIDAR DATA PROCESSING
Z. Lari, A. Habib
Dept. of Geomatic Engineering, University of Calgary, 2500 University Drive N.W., Calgary, AB, T2N1N4, Canada
(zlari, ahabib) @ucalgary.ca
Commission III, WG II/2
KEY WORDS: LIDAR, Point cloud, Processing, Estimation, Quality, Analysis
ABSTRACT:
Over the past few years, LiDAR systems have been established as a leading technology for the acquisition of high density point
clouds over physical surfaces. These point clouds will be processed for the extraction of geo-spatial information. Local point density
is one of the most important properties of the point cloud that highly affects the performance of data processing techniques and the
quality of extracted information from these data. Therefore, it is necessary to define a standard methodology for the estimation of
local point density indices to be considered for the precise processing of LiDAR data. Current definitions of local point density
indices, which only consider the 2D neighbourhood of individual points, are not appropriate for 3D LiDAR data and cannot be
applied for laser scans from different platforms. In order to resolve the drawbacks of these methods, this paper proposes several
approaches for the estimation of the local point density index which take the 3D relationship among the points and the physical
properties of the surfaces they belong to into account. In the simplest approach, an approximate value of the local point density for
each point is defined while considering the 3D relationship among the points. In the other approaches, the local point density is
estimated by considering the 3D neighbourhood of the point in question and the physical properties of the surface which encloses
this point. The physical properties of the surfaces enclosing the LiDAR points are assessed through eigen-value analysis of the 3D
neighbourhood of individual points and adaptive cylinder methods. This paper will discuss these approaches and highlight their
impact on various LiDAR data processing activities (i.e., neighbourhood definition, region growing, segmentation, boundary
detection, and classification). Experimental results from airborne and terrestrial LIDAR data verify the efficacy of considering local
point density variation for precise LiDAR data processing.
1. INTRODUCTION
In recent years, Light Detection and Ranging (LiDAR) systems
are being extensively utilized for rapid collection of high
density 3D point clouds. These point clouds provide accurate
3D data for different applications such as terrain mapping
(Elmqvist et al., 2001), transportation planning (Uddin and Al-
Turk, 2001), 3D city modeling (Kim et al., 2008), heritage
documentation (Patias et al. 2008), and forest parameter
estimation (Danilin et al., 2004). The collected data should be
processed to extract useful information for the aforementioned
applications. The internal properties of LiDAR point cloud and
the performance of processing techniques highly affect the
validity of extracted information. Point density is an important
characteristic of LiDAR data that should be considered during
the various processing activities (e.g., neighbourhood
definition, classification, segmentation, feature extraction, and
object recognition).
The majority of existing data processing techniques assumes
that the LIDAR point cloud has a uniform point density.
However, the collected data might show variations in the point
density due to irregular movements of the acquisition platform,
variations in the scattering properties of the mapped surface,
number of overlapping strips, and incorporation of terrestrial
laser scanning systems (Vosselman and Mass, 2010). Therefore,
data processing techniques should consider possible variations
in the point density within the datasets in question. This ability
will assure the flexibility of these techniques in dealing with
datasets captured by different platforms and/or from different
sources. The effective consideration of these variations is
125
dependent on providing standard approaches for the estimation
of local point density indices.
So far, only a few applicable methods have been presented for
the determination of local point density indices. The most
commonly used method for the estimation of local point density
is the “box-counting method” proposed by County (2003). In
this method, the LiDAR points are firstly projected onto a 2D
space. Then, the defined 2D space is rasterized using a grid with
a predefined cell size. The local point density index for the
points within each cell is then determined as the total number of
the projected points in that cell normalized by the cell area. The
main disadvantage of this method is that the estimated local
point density values are dependent on the size of the cells and
their placement within the projected data. Since there is no
standard for the determination of the appropriate cell size and
its placement within a LiDAR data, different density values may
be estimated for the same points in a LiDAR dataset (Raber et
al., 2007).
In order to estimate unique local point values for individual
LiDAR points, Shih and Huang (2006) proposed a TIN-based
method. In this method, the point cloud is firstly triangulated
using an advanced Delaunay triangulation technique (Isenburg,
2006) to produce a TIN. Then, a Voronoi Diagram is
constructed utilizing the generated TIN structure. The local
point density index for each point is then estimated by inversing
the area of the Voronoi polygon assigned to that point. The
presented approaches for the local point density estimation only
consider the 2D distribution of the point cloud. Therefore, they
are only valid when dealing with point cloud acquired by an
