X-B3, 2012
cient RANSAC
araphies Forum
Photo Tourism:
Pransactions on
ua, P. and Thi-
era Calibration
Imagery. In
attern Recogni-
esh Segmenta-
egion Growing.
E-792,
Processing. Im
46,
amd Structuring
ser Scanning,
| Detailed Digi-
ratnted Stmpli-
3ecinformation
07. Detection
ectified Facade
3eoinformation
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
PROGRESSIVE DENSIFICATION AND REGION GROWING METHODS FOR LIDAR
DATA CLASSIFICATION
J. L. Pérez-García *, J. Delgado *, J. Cardenal ^, C. Colomo “, M. A. Ureña *
* Dpto. Ingeniería Cartográfica, Geodésica y Fotogrametría. Escuela Politécnica Superior. Universidad de Jaén. Campus
Las Lagunillas s/n. 23071 Jaén (Spaín).
Email: jlperez@ujaen.es, jdelgado@ujaen.cs, jcardena@ujaen.es, emej0002@estudiante.ujaen.es, maurena@ujaen.es
Commission III, WG II[/2
KEY WORDS: LiDAR, DEM/DTM, Classification, Algorithms
ABSTRACT:
At present, airborne laser scanner systems are one of the most frequent methods used to obtain digital terrain elevation models.
While having the advantage of direct measurement on the object, the point cloud obtained has the need for classification of their
points according to its belonging to the ground. This need for classification of raw data has led to appearance of multiple filters
focused LiDAR classification information. According this approach, this paper presents a classification method that combines
LiDAR data segmentation techniques and progressive densification to carry out the location of the points belonging to the ground.
The proposed methodology is tested on several datasets with different terrain characteristics and data availability. In all case, we
analyze the advantages and disadvantages that have been obtained compared with the individual techniques application and, in a
special way, the benefits derived from the integration of both classification techniques. In order to provide a more comprehensive
quality control of the classification process, the obtained results have been compared with the derived from a manual procedure,
which is used as reference classification. The results are also compared with other automatic classification methodologies included in
some commercial software packages, highly contrasted by users for LiDAR data treatment.
1. INTRODUCTION can be used for removing non-ground objects from the DSM
data, such as, building and trees.
In recent years, there has been a significant increase of the Many authors have proposed different filtering methods based
widespread use of airborne laser scanning systems for obtaining on this type of operators. Linderberger (1993) proposed the first
digital elevation models with high-resolution spatial accuracy. filter based in the use of mathematical morphology applied to
The key advantage of these systems is that the point ^ the filtering of profiles captured by a airborne laser profiler,
measurement is made in a direct manner using a laser ^ using a opening process and considering a structural element
distanciometry system. However, these systems have the with constant size and a certain threshold. Vosselman (2000)
problem of not carrying out any interpretation of the scene, so introduces an distance-dependent threshold (slope) with a
the points are distributed on the ground according to certain constant size structural element. This modification implies a
pattern depending of the scanning instrument of the LiDAR, the larger tolerance for more distant points.
flight and data capture configurations and the terrain
characteristics. This implies, in the case of the applications According the same approach, Sithole (2001) and Roggero
where the desired product is the ground surface, the need for a (2001) incorporated the local terrain slope at each point to the
subsequent filtering process for classifying data as belonging to structural element definition instead of working with an average
the ground or belonging to the different objects located on it. slope for the entire area (Vosselman filter). Thus, these new
filtering methods have structural elements with constant size but
1.1 LiDAR data filtering methodologies for DTM variable form. Kobler (2007) proposed the structural element
generation. rotation in order to adapt it to the local slope. Considering that
the use of a fixed-size structural element can provide not
The interest on LiDAR data filtering methodologies has satisfactory results, several authors use variable sizes for the
supposed the apparition of different algorithms and methods in structural element, instead the fixed sized of the above-
order to differentiate between those the points belonging to the mentioned methods. For example, Kilian et al. (1996) perform
ground, from those not belonging to it. These methods have various opening processes with different window sizes. They
different approaches and usually they are classified into five use the window size as weight in the calculation of the final
different groups. terrain height. Lohmann et al. (2000) proposed the Dual-Rank
filter, which first uses a morphological filter in order to remove
the non-ground points before carrying out the calculation of the
final model. Zhang et al. (2007) proposed a progressive
Using mathematical morphology operators, such as erosion and ^ morphological filter based in the progressive enlargement of the
dilation filtering procedures implemented in LiDAR data, are window size. Using similar approach, Arefi and Hahn (2005)
characterized by the use of a structural element. These methods propose a progressive morphological filter using a geodetic
distance operator.
1.1.1 Morphological filtering.
131
