You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Close-range imaging, long-range vision

sw variation
RPA Image
Zurich city
lose range
ium of the
Single view
Jan Bôhm
Institute for Photogrammetry, University of Stuttgart, Geschwister-Scholl-Str. 24, 70174 Stuttgart, Germany -
KEY WORDS: Segmentation, Classification, Extraction, Surface Feature, CAD
This paper addresses the problem of segmenting dense range data containing curved surfaces. Segmentation is a crucial
step in the processing of range data for applications in object recognition, measurement, reengineering and modeling.
We propose a two stage process using model-based curvature classification as an initial grouping. Features based on
differential geometry, mainly curvature features, are ideally suited for processing objects of arbitrary shape including of
course curved surfaces. The second stage uses a modified region growing algorithm to perform the final segmentation.
The approach is demonstrated on a test scene acquired with a stripe projection sensor.
Today a full variety of range sensing devices is available
for all kinds of object sizes, resolution demands and pur-
poses. Applications range from industrial inspection, multi-
media, as-built-documentation to cultural heritage and oth-
ers. Be it active triangulation or time-of-flight systems,
scanners are able to densely sample surfaces with great ac-
curacy. Oftentimes hundreds of thousands or even millions
of points are obtained from only a single scan. Usually the
result of a single scan is represented as an ordered point
cloud. Most of the times the topology is based on a grid
either given by a frame sensor or by the scan movement.
While this type of data is easily triangulated and directly
suited for visualization purposes, most applications require
data of another quality. Be it for measurement purposes,
reengineering, construction or design purposes it is nec-
essary to give an interpretation to the point cloud, that is
to group individual points into meaningful entities. This
grouping is called a segmentation. Just what ’meaningful’
really is, is determined by the application. In the context
of range data, most often a segmentation into individual
surfaces is desired.
Several possibilities exist for the segmentation of dense
range data. Among the first and most prominent are region
growing approaches. The crucial point in region growing
is the selection of seed regions. When seed regions are se-
lected to close to the rim of a surface, the resulting region
is usually meaningless. Today we see the first implemen-
tations of region growing approaches in commercial appli-
cations. The problem of seed region selection is left to the
operator, who selects regions manually. This manual oper-
ation can be quite time consuming considering the size of
the point cloud and is generally a tedious task. Thus it is
evidently desirable to automate the process.
The work by J.P. Besl (Besl, 1988) has been among the first
to present automated region growing for range image seg-
mentation containing curved surfaces. By initially group-
ing the pixels of a range image according to the sign of
the Gaussian and Mean curvature an over segmentation is
achieved. By successively applying morphological opera-
tors the region are shrunk to minimal size. These minimal
regions are then used as seed regions for region growing.
While the algorithms were shown to produce good results,
the complexity of the approach and the large number of
user supplied parameters have prevented the wide-spread
adoption of the approach in practice.
In the course of our work on industrial inspection using
range imaging, we have developed methods for range im-
age processing and feature extraction. We have earlier re-
ported on our developments of a range image classifica-
tion framework using differential geometry (Bóhm et al.,
2000). Features based on differential geometry, mainly
curvature features, are ideally suited for processing objects
of arbitrary shape including of course curved surfaces. Fur-
thermore these features are suited for object recognition
purposes as they are invariant to rotation and translation.
While the curvature-based classification approach gives an
interpretation to each individual pixel and by performing a
cluster analysis in curvature space effectively gives a group-
ing of the pixels, this grouping does not take into account
the topology of the point cloud. This is a major shortcom-
ing when comparing the classification result to a segmen-
tation. Using the classification result as an initial grouping
we now present our work on extending the classification to
a segmentation of the point cloud. This two stage process
uses a modified region growing algorithm to perform the
segmentation based on the topology. While our approach
shares some concepts with the one mentioned above, it ex-
tends the idea in several ways. First the absolute value of
curvature is used for initial classification, not only the sign
of curvature. Secondly model knowledge of the object is
incorporated into the process and thus greatly enhances re-
liability. When the type of surface can already be derived
from model knowledge, this information can be used to
check for the homogeneity of a group of points during the
growing process.
While the proposed framework in principle is general to
range data, we show its application in the context of indus-
trial measurement purposes. All range data for this project
has been acquired with a stripe projection system devel-
oped at the Institute for Photogrammetry. The scanner sys-
tem itself is described in detail in the next chapter including