International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
  
  
  
  
    
  
  
seed 1 «- seed 2 
tower 1 building | —_ ; D 
pm E 
| L C2 = 
© seed 4 . — seed 3 
grund free! res 2 c2 pz 
  
  
  
primitive (p) and complex (c) 
contours objects 
tower 1 tower 2 
n 5 tree 1 tree 2 
A = (child, p) 
building (c1) |! F3 
n ground (root/parent, c2) 
Figure 4. Object hierarchy 
In the next step, the object family tree is analyzed from the 
leaves (the top/seed of the objects) to the root to determine, if 
an object is a real-world object sitting on the ground or if an 
object is part of a terrain structure (or an artificial structure that 
should be considered terrain like a dam). This is done by 
analyzing the shape of the object and the “growing behavior” of 
its segments from level to level. In a simple case, a seed may be 
a small spot that grows quadratically in area from level to level 
into a rectangular shape, stops growing for several levels and 
suddenly grows again in large but random steps. The real-world 
object described is a simple house with a rectangular floor plan. 
It is “completed” on the last (lowest) level before the segment 
starts growing randomly. 
Numerous criteria can be used to determine real-world objects 
and on which elevation level they start: object geometry (area 
growth, contour/segment shape, relationship of area size and 
contour length, relationship of volume (height) and area size, 
etc.), object attributes, object context (shapes, sizes, growth 
behavior of "sibling" (adjacent) objects, and the parents and 
grandparents of multiple objects in the object tree), and several 
others. How a real-world object is detected from these criteria is 
controlled by a fuzzy-logic-based classification that may be 
parameterized by training. 
This approach has proven to work very well with a large variety 
of surfaces including mountainous areas with steep slopes and 
sharp ridges that often pose challenges to conventional 
approaches. We also found that discrimination of artificial 
structures like dams and ramps that are usually considered as 
belonging to the “ground” class are reliably classified as such 
while other artificial structures of almost any size and shape are 
reliably identified as non-ground (e.g. buildings). 
Besides distinguishing surface objects from ground, contour- 
based object detection generates comprehensive information 
about the surface objects that can readily be used for further 
classification. Different types of surface objects (buildings, 
trees, etc.) can be identified and geometrical object descriptions 
(building geometry, roof shape, ridge orientation etc.) can be 
836 
derived with little additional effort by just evaluating the 
geometry of the abstract object. Complex buildings, for 
example, are represented in the object tree as a parent object 
with multiple child objects representing building primitives (see 
figure 4). The geometries of parent and child objects structure 
the building and describe its geometry in detail. 
In the other direction a generalized surface and land use 
description may be derived by classifying lower-level complex 
objects according to their children (e.g. an area with many 
small-sized child objects of type “building” makes a housing 
area, similarly: industrial areas, parks, forest, agricultural areas 
cic.) 
It should be mentioned that this approach is easily expanded to 
use attribute data like in form of 3D-points and raster images as 
well as any available a-priori information in object or structural 
from, and also including return waveform data, to increase the 
selectivity and robustness of the classification and to refine the 
object descriptions: 
=  3D-point data within.object boundaries, point density and 
the planimetric and vertical distributions of points within 
an object indicate its "transparency", 
* lidar intensity, surface color, and raster image data provide 
radiometric information about an object, 
*  vectorized object descriptions from digital catastral maps, 
zone delineations, breaklines, or segmentation maps from 
digital imagery, etc. can be used as additional input to the 
segmentation process to refine the output or to define 
objects that cannot easily be recognized from geometry 
alone (administrative zones, for example) 
* return waveform data describes surface attributes like 
“roughness” and “slope”, and, representing reflectance 
within the volume (at each level inside an object), the 
content and internal structure of transparent and partially 
transparent objects. 
3.4 Interactive Inspection and Refinement 
Depending on the requirements to the quality of the 
classification LasTools contains an easy-to-use, flexible, and 
fast data viewer and editor for interactive verification and local 
refinement of the classification results. 
Data Viewer 
The data viewer enables the user to visualize all relevant 
aspects of the available data. It displays 
= vector data (outlines of regions, blocks, strips of lidar data, 
planned flight line vectors and flown trajectories, imported 
breaklines and vector maps, profiles, contour lines, 1D and 
2D histograms ...), 
" point data (3D lidar measurements) colored by attributes 
(timestamp, line number, return number, elevation, spot 
height, intensity, surface color, …) 
= raster data (DTM/DSM color-coded elevation, shaded 
relief, intensity and surface color images, digital 
orthoimages, imported raster maps), 
= volumetric data (volume reflectance from waveform data, 
slices, profiles, etc.) 
simultaneously in a data (world) coordinate system. Multiple 
views may be synchronized to show different aspects of a data 
set in separate windows that are updated as the user pans or 
zooms one of them. 
The user may select and manipulate the display area (pan, 
zoom, rotate), switch between data layers, superimpose and mix 
Di 
pc 
3l 
ha 
pr 
da 
an 
in 
the 
sel 
us 
gk 
mi 
CO 
fui