CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
66
equipment was precisely installed by topometry. It allows the
gathering of georeferenced 3D laser data with a very high density
and also much information about the acquisition (see section 4).
Figure 1: Visualization of the 3D point cloud of very high den
sity. This raw data represents the building facade acquisition. The
black line represents the trajectory of the laser sensor.
The laser data are acquired under realistic conditions in dense
urban environments. Moreover, the datasets are collected in a
particular container related to the laser scanner, particularly in a
range of 3D points. In this context, the data must be manipulated
with much precaution. Here we will describe the main difficulties
associated with 3D data.
• The acquisition: The laser sensor data can be represented
in two ways; either as a container with data organized lin
early in temporal sequences of 201 points (data frames) or
like a cloud of 3D points (georeferenced data). The condi
tions of acquisition are very variable in realistic and dense
urban environments. Mobile objects cause a thickening of
the acquired cloud when the vehicle stops. Certain acquired
points model an ephemeral surface and could be considered
as erroneous points. Moreover, the density of the facades
vary according to the speed of the vehicle.
• The occlusions: pose a problem for the complete acquisi
tion of building facades in urban environments. The occlu
sions could be caused by two categories of obstacles, static
or dynamic created by man-made and natural objects. The
raw cloud may suffer from missing data due to the pres
ence of pedestrians, trees, mobile and parked vehicles and
many others objects (see figure 2). Alas, this inevitable phe
nomenon affects the modeling process.
• The laser reflectance: could cause confusions in the 3D
data interpretation. Certain points don’t model a physical
surface. This effect appears on a retroreflector surface. Ob
servations sometimes show an aureole of points around road
signs. These dispatched points represent erroneous data.
Moreover, certain points model a different surface other than
the surface of interest. Sometimes, the beam of the laser ei
ther rebounds off of the outside of the window or it passes
through the window and models the inside of the dwelling.
These scattered points represent erroneous information for
the facades modeling. In addition to this, other less frequent
effects could arise due to poorly reflective surfaces. •
• The redundancy of data: is due to many factors. The ac
quisition is continuous even when the vehicle is stopped.
We have adopted this strategy to facilitate the acquisition of
a large area and to use data as common bases for our dif
ferent projects. Therefore, raw laser data may contain many
redundant frames. Moreover, due to sensor characteristics
(orientation and linear scanning), we could sometimes have
up to three acquisitions of the same facade part caused by
the graining of the laser beam in the turns. The redundancy
of data (points, frames, parts of the facade) presents an in
convenience for the feature extraction techniques based on
vote schemes or random trials.
Figure 2: A street in the city of Paris. The building facade is
partially occluded by trees and parked vehicles.
Figure 3: Returned intensities of the 2D scans. The redundancy
effect appears when the vehicle temporarily stops. The vehicle
and the branches seem stretched.
This brief description allows us to acknowledge several problems
associated with the raw laser data. The 3D data should undergo
several preprocessing steps before becoming exploitable. Thus
we need a process robust to some outliers and noisy data.
3 PROPOSED APPROACH
In this section, we describe our approach which consists of two
stages. The first stage focuses on the 3D cloud points preprocess
ing. The second stage aims at the building footprint extraction.
In this work, we assume that buildings have simple polygonal
shapes.
3.1 3D data preprocessing
3.1.1 Partial Altering of redundant points As we have men
tioned earlier, the laser sensor constantly sweeps the building fa
cade even when the vehicle is stopped. Consequently, the ac
quired raw data may contain many redundant frames due to this
continuous acquisition. For this reason, we have defined a mea
sure between two consecutive frames based on point-to-point dis
tances. The redundant frames are thus detected and removed from