International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
AN ITERATIVE TERRAIN RECOVERY APPROACH TO AUTOMATED DTM
GENERATION FROM AIRBORNE LIDAR POINT CLOUDS
Hufeng Chen *, Ming Cheng * Jonathan Li ** *, Yinsheng Liu °
? GeoSTARS Lab, Department of Geography and Environmental Management, University of Waterloo, 200 University
Ave. West, Waterloo, ON, Canada N2L 3G1 - (h96chen, junli)@uwaterloo.ca
? GeoSTARS Group, School of Information Science and Engineering, Xiamen University, 422 Siming Road South,
Xiamen, Fujian, China 361005 - chm99@xmu.edu.cn
* Hunan Provincial Transport Technical Information Center, Changsha, Hunan, China
KEY WORDS: DTM, LiDAR, Terrain filtering, Multi-scale, Pyramid level
ABSTRACT:
This paper presents a hierarchical recovery method to generate DTMs from airborne LiDAR point clouds based one an idea of
layering. The developed method first registers the last return points, and then layering them. The layering is done by dividing the
points into different height layers and assigning layer numbers to each point. The layer numbers are comparing references in later
identification process. Then a series of rasterized pyramid levels which consisted of lowest points are generated. Since the outliers
have been removed after the layering, the cells in top level are considered to be terrain points and used as reference to identify cells
in the following level. After the identification of the second level, an interpolation will occur in the cells which identified as off-
terrain. And the interpolated level will be used as reference in its following level and the same process is repeated at each level. Once
this process of the bottom level finished, the proposed method adjusts the results based on the first return feedback and followed by
the final interpolation. As a result, this produces the final DTM. The developed method is data driven, and does not assume a prior
knowledge about the scene complexity. The proposed method was tested with the ISPRS WG III/3 LiDAR datasets covering
different terrain types and filtering difficulties. The results show that the proposed method can perform well in flat terrain or gentle
slope area.
1. INTRODUCTION
Compared to conventional methods such as aerial
photogrammetry and field surveys, the generation of Digital
Terrain Models (DTM) from airborne LiDAR point clouds is
fast and cost-effective over a large area, especially in vegetation
covered areas since laser pulses can penetrate some of the
canopy. However , developing an automated and robust
approach to terrain point identification and DTM generation is
challenging. As a preliminary task of DTM generation using
airborne LiDAR point cloud data, filtering terrain and off-
terrain is critical and fundamental to feature extraction and
classification (Briese, 2010). The identified terrain points are
the input of further interpolation process in many developed
algorithms. The inappropriate identification will cause the
deviation in the following interpolation, which leads to further
error and low accuracy of DTM products (Guo et al., 2010).
Besides, filtering is usually very challenging and time
consuming because of the algorithms have to processing a large
amount of data. Therefore, an efficient and effective filter
algorithm is important for DTM generation.
However, the current filtering algorithms are facing difficulties
in handling complex circumstances such as outliers (points lie
far above or below the most points), complex objects, steep
slopes, attached objects, uncertainty of the terrain definition
(such as the ramp of a bridge), vegetation (such as shrub),
discontinuities of the terrain, low objects like curb and railway
tracks, as well as the combined complex scene (Sithole and
Vosselman, 2004; Meng et al., 2010). Some of these problems
are critical. The outliers, especially low outliers, can affect the
reference point selection in the algorithms which adopt the
lowest points as reference terrain points. The scenarios with
* Corresponding author. Phone: 1 519-888-4567, ext. 34504.
different sizes of the buildings will face a dilemma in choosing
a filtering window size. Applying a small window size, the
algorithm may identify a point on a big building point as terrain;
and using a large window size, small terrain relief variations
may be ignored. Low elevation objects are hard to remove
because their heights are very close to terrain. A lot of research
has been dedicated to DTM generation, especially to filtering,
during the last decade (Meng et al, 2010). However the
aforementioned difficulties have always been a barrier in
developing a fast, robust, and reliable automatic filter and it is
the major obstacle in DTM generation from airborne LiDAR
data.
The rest of the paper is organized as follows. Section 2
describes the multi-scale terrain filtering method. Section 3
discusses the results obtained using the ISPRS WG III/3 test
LiDAR datasets. Finally, Section 4 concludes the study.
2. METHOD
The proposed multi-scale terrain filtering (MTF) method
identify terrain points by iteratively recover terrain models from
multi-scale pseudo-grid images in a coarse-to-fine way. AS
shown in Figure 1, the method has three steps: point cloud pre-
processing, multi-scale terrain filtering, and DTM refinement.
In the point cloud pre-processing step, all laser scanning points
must be pre-processed to retain last-return points of multiple
returns (laser echoes), and then are layered with regard to the
statistical height histogram of the whole data set. Two
objectives of height histogram based layering are to assign the
layer numbers to each point for the following MTF
implementation, and to remove lower outliers and noises.
Input Li
//Pre-pr
points.S
points.H
points.L
points.N
//Points
level[n] :
for(level
if(j-
for(
for(
//DTM R
Rough L
Refined
Lida
Terrain I
Lida
FinalDT^
Output D
——
Figui
The secon
Includes the
identificatio
levels are g
every level
level is refe
the propos
identificatio
