CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
14
2 METHODOLOGY
Building and road extraction, as mentioned above, require com
plex elaborations of multi-source data; we followed a multi-step
procedure. The procedure here proposed consists of four sequen
tial steps; the output of each module is the input for the following.
Step 1 - Feature generation. It calculates LiDAR and radio-
metric additional features for the classification stage; a total
of seven mixed-features are currently adopted.
Step 2 - Classification. Using AdaBoost with a tree classifer
as weak learner, it distinguishes among four main classes; a
simple training set is adopted to train the classifer.
Step 3 - Object Extraction. It extracts buildings and/or roads
from the classified data; in this paper we focus on road ex
traction and pre-filtering techniques;
Step 4 - Clustering. It is fundamental to model the extracted
objects.
A graphical representation of discussed methodology is shown in
Fig-1-
Figure 1: Methodology. The object extraction procedure has a hi
erarchical structure that simplifies the phase of result evaluation;
different approaches can be easily tested without compromising
the overall methodology
In the following sections, the results of each stage are presented;
for completeness a deep results evaluation of building extraction
is reported to evidence the quality of classification process; stan
dard metrics are used to make in evidence the performance of
AdaBoost classifier.
3 CLASSIFICATION
3.1 Dataset
The methodology presented in previous section, was validated in
an urban area: LiDAR and multi-spectral data refer to the centre
of the German city of Mannheim. This area is characterized with
large buildings, mostly attached forming building blocks of dif
ferent heights, many cars and little vegetation. Mannheim dataset
has a resolution of 0.25m for the images and 0.5m for the range
data; the total grid dimension is 1808 x 1452 (width x height).
The aerial images are orthorectified and four spectral bands are
available: Red, Green, Blue, and Near InfraRed; laser range data
consist of first and last pulse recordings acquired by an airborne
laser scanner. Additional features were added to expand the fea
ture space; main motivation is that using a feature weighting al
gorithm, is easy to find the best feature combination. Normal
ized Difference Vegetation Index (NDVI) and Green Normalized
Difference Vegetation Index (GNDVI) were calculated. These
indexes are useful to distinguish between some critical classes
which LiDAR data cannot easily distinguish. Two pairs are criti
cal: building/tree and land/grass. NDVI is a compact index which
allows to better discriminate inside each cited pair. It is well
known that canopies and grass have a NDVI value usually greater
than 0.15, while for building and land classes is usually around or
below zero. As introduced in the previous sections, we identified
four main classes; for each class, we selected eight representative
polygons. The total area of training set is below the 0.5%; it is
useful to remark that the selection of these polygons is a low-time
consuming activity that can be easily performed using a web-GIS
or photo-interpretation (easy owing to the reduced number and
kind of classes). The training set and a 3D view of the input data
set are shown in Figures 2 and 3.
Figure 2: Data and Training set. Red stands for building, yellow
for land, blue for grass and green for tree
Figure 3: A 3D view of dataset; height of objects are obtained
using the first pulse laser range data
The selected features used for classification are:
LiDAR: Ah is the height difference between the last pulse DSM
and the DTM and Ap is the height difference between the
first pulse and the last pulse DSM
Spectrals: R,G,B,NIR and NDVI (GNDVI is omitted because
the weight associated to this feature was low)
