745
SEGMENTATION OF TERRESTRIAL LASER SCANNING DATA
BY INTEGRATING RANGE AND IMAGE CONTENT
Shahar Bamea, Sagi Filin
Transportation and Geo-Information, Civil and Environmental Engineering Faculty, Technion - Israel Institute of
Technology, Haifa, 32000, Israel - {bameas, filin}@technion.ac.il
Commission V, WG V/5
KEY WORDS: Segmentation, Terrestrial Laser Scanner, Point Cloud, Algorithms, Object Recognition
ABSTRACT:
Terrestrial laser scanning is becoming a standard technology for 3D modeling of complex scenes. Laser scans contain detailed
geometric information, but still require interpretation of the data for making it useable for mapping purposes. A fundamental step in
the transformation of the data into objects involves their segmentation into consistent units. These units should follow some
predefined rules, and result in salient regions guided by the desire that the individual segments represent object or object-parts within
the scene. Nonetheless, due to the scene complexity and the variety of objects in it, it is clear that a segmentation using only a single
cue will not suffice. Considering the availability of additional data sources like the color channels, more information can be
integrated in the data partitioning process and ultimately into the reconstruction scheme. We propose in this paper the segmentation
of terrestrial laser scanning data by the integration of range and color content and by using multiple cues. This concept raises
questions regarding the mode of their integration, and definition of the expected outcome. We show, that while individual
segmentation based on given cues have their own limitations; their integration provide a more coherent partitioning that has better
potential for further processing.
1. INTRODUCTION
Terrestrial laser scanners emerged in recent years as standard
measuring technology for detailed 3D modeling of scenes. From
a geometrical perspective, scanners provide rich and accurate
information of the acquired scene. Additionally, with cameras
becoming an integral part of modem scanners, the resulting
radiometric information provides supplementary color content.
The combination of direct geometric details and radiometric
content offers excellent foundations for the extraction of objects
in an autonomous manner.
Raw data (3D points and 2D RGB pixels) resulting from a
single scan can reach tens of millions of elemental units.
However, for common laser scanning applications, e.g.,
mapping, modeling, and object extraction that require high level
of abstraction, this huge amount of data is hard to use. A
fundamental step in the extraction of objects is the application a
mid-level processing phase involving the grouping of pixels
containing redundant information into segments. Essentially,
each segment should form a collection of 3D points in which
two conditions must be met, one is that the segment will
maintain geometrical connectivity among all points constituting
it; the second is that the feature value for the connected points
will share similarity of some measure. Similarity can be
geometrically based, radiometric based, or both. In addition, the
basic units of each segment have to create a spatial continuation
in the 3D sense. While segmentation of image content, and to
some degree, of terrestrial point clouds, has been studied in the
past, segmentation of the combined set has not been addressed
by many so far. The motivation for pursuing this avenue is
however clear and relates to the desire to benefit from the
descriptive power of the rich radiometric content while being
subjected to objects geometry and spatial connectivity in 3D
space.
In general, segmentation concerns partitioning the data into
disjoint salient regions usually under the assumption that
individual segments tend to represent individual objects within
the scene. Due to its important role, segmentation has been
studied for years beginning from thresholding techniques (Otsu,
1979; Huang et al., 2005) and classic "region growing" based
methods (e.g., Pal and Pal, 1993). Other methods propose
converting the image into a feature space, and by doing so
transforming the segmentation problem into a classification task.
Carson et al. (2002) propose modeling the distribution of feature
vectors as a mixture of Gaussians, with the model parameters
being estimated using the expectation-maximization algorithm.
Graph based approaches have been receiving growing attention.
Using this scheme images are viewed as a graph in which each
vertex represent a pixel (Shi and Malik, 2000; Felzenszwalb and
Huttenlocher, 2004). The graph-view enables an intuitive
representation of the segmentation problem as similarity
between pixels can be assigned to the edges linking them. The
challenge is then to find sets of vertices such that each has high
connectivity value between its vertices and low connectivity to
the rest of the graph. For a computational model for such
segmentation, normalized cuts algorithm has been proposed
(Shi and Malik, 2000). Sharon et al. (2000) make use of the
multi-grid theory (Brandt, 1986) to solve efficiently the
normalized-cut problem. A comprehensive review and test of
some of the leading segmentation algorithms is provided in
Estrada and Jepson (2005). Recent works, e.g., Russell et al.
(2006), Roth and Ommer, (2006), Mian et al. (2006), and Alpert
et al., (2007) demonstrated the application of segmentation
processes for recognition tasks, showing promising results both
in relation to object class recognition and to correct
segmentation of the searched objects. Applications making use
of segmentation as part of other tasks, have been reported for
stereovision and image registration purposes (Bleyer and
Gelautz, 2004; Klaus et al., 2006; Coiras et al., 2000).