ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002
introduce unnecessary distortions. So, in addition to the
algorithmic concerns, an adaptation of image-processing
concepts to the irregular pattern becomes necessary. With
the aim of identifying structure in the data in mind, the pro-
posed algorithm is general and can be applied in a variety
of applications.
2 SURFACE CLUSTERING
Similar to data segmentation, the goal of data clustering
is to subdivide the data into disjoint regions each with a
homogeneous property that distinguishes it from its sur-
rounding. The regions are defined by the set of points in-
cluded within each segment, where different regions can-
not share points. The proposed algorithm considers the
surface clusters an instantiations of more generic processes
defined here as surface categories. The algorithm aims at
distinguishing among four different surface categories, i)
forested/wooded area, ii) low vegetation areas and rough
surfaces iii) smoothly varying topography, and iv) planar
surfaces. Surfaces refer here to the interpretation of the
data obtained by laser scanning system, and the categories
present one interpretation of the laser data surface. The
surface classes are not aimed at providing a topographic
structure of the terrain, mainly since the acquired data is
not the terrain itself. Yet the clusters provide a separation
of the surface into homogeneous parts. The distinction be-
tween smoothly varying topography and planar surfaces is
made here because of the tendency of man-made object
to have planar facets, and the value of this information to
other applications.
2.1 The feature vector
By nature, laser data attributes will be derived from surface
texture measurement.! The measures should be sufficient
to differentiate among surface categories and among sur-
faces within each category. Several measures have been
proposed for segmentation of range data, among them are
the analysis of the height differences in a window via his-
togram followed by segmentation based on thresholding.
Axelsson (1999) uses the second derivatives to find varia-
tions, and Maas (1999) uses a feature vector including the
Laplace operator, maximum slope measures and the origi-
nal height data in order to classify the data.
In this implementation clustering is performed based on
an attribute vector consisting of the following measures —
the point position, the parameters of the tangent plane to
that point, and the relative height difference between the
point and its neighbors. Together they form a 7-tuple vec-
tor vj :— (zi, yi, Zi Tí 1,230 pi, di] for each point, with
Ti, Yi, Zi, the laser point coordinates, 7/, »3;, Pi, the sur-
face parameters (normal measured by two parameters and
a constant), and d; the height difference of the point to its
neighbors.
The inclusion of the point position as an attribute is es-
sential for measuring proximity to other points that share
1]f reflectance measurements are available, information can also be
derived from these values.
A - 120
Figure 1: Potential inseparability of surfaces based on
height differences
similar properties. Height differences are perhaps the most
commonly used measures as they measure local variation
and are expected to be reliable up to the level of noise in
the data. They provide an adequate indication for the exis-
tence of high vegetation but they are insufficient for surface
separation as the example in Figure 1 demonstrates. From
an analytical standpoint height differences capture the ex-
istence of step edges in the data, and emulate the effect of
an edge operator in raster data. Their main contribution is
in enhancing the separation of clusters from one another.
The tangent plane parameters consist of the normal direc-
tion and the constant. The slope parameters capture first-
order discontinuities, thereby enhancing the separation of
surface elements with different trend such as the ones in
Figure 1. Slopes capture no positional information, but the
constant value positions the plane in space and enables sep-
arating surfaces with similar slopes. Surface parameters
and the height differences share some similarity; consider
for example two horizontal planes for which the difference
in the surface constant is in-fact the height difference. Yet,
the plane constant refers to an infinite plane and is a rather
global measure while the latter measures difference to a
neighboring points, and is rather local.
Category Surface Height
Slopes Difference
High vegetation | Rapidly varying Large
Low vegetation | Rapidly varying | Medium
Smooth surface | Locally constant Small
Planar surface Fixed Small
Table 1: Surface categories vs. attributes
Table 1 lists the expected characteristics of these attributes
for each of the four surface categories. The measures are
qualitative and not strict, but they are sufficient to indicate
that the chosen features make the identification of these
four categories possible. Translating these measures into
quantitative values is partially the essence of the algorithm.
2.2 Metrics to measure the attributes
Successful clustering of the data depends on the features
representation. While height differences consist of a sin-
gle measure and have a natural metric unit, planar surfaces
can be described in various ways. It is common to use
the explicit three parameters representation consisting of
the slopes in the x and y direction, s,, s, and the intercept
point. However, this representation breaks down with ver-
tical or near vertical structures (e.g., walls.) The represen-
tation of the surface slopes by their tangents also offers