3HHi
In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part 3A - Saint-Mandé, France. September 1-3. 2010
281
ROAD EXTRACTION AND ENVIRONMENT INTERPRETATION FROM LIDAR
SENSORS
Laurent Smadja, Jérôme Ninot and Thomas Gavrilovic
VIAMETRIS. Maison de la Technopole, 6 rue Leonard de Vinci, BP0119, 53001 Laval cedex, France, www.viametris.fr
Commission III, WG 111/2
KEY WORDS: LIDAR / Camera calibration, Unsupervised road extraction, Interpreted 3D reconstructions
ABSTRACT:
We present in this article a new vehicle dedicated to road surveying, equipped with a highly precise positioning system. 2D lidar scans
and high definition color images. We focus at first on the sensors extrinsic calibration process. Once all sensors have been positioned
in the same coordinates system, 3D realistic environments can be computed and interpreted. Moreover, an original algorithm for road
extraction has been developed. This two-step method is based on the local road shape and does not rely on the presence of curbs or
guardrails. Different uses of the RanSaC algorithm are employed, for road sides rough estimation in the first place, then for unlikely
candidates elimination. Road boundary and center points are further processed for road width and curvature computation in order to
feed a geographic information system. Finally, a simple extraction of traffic signs and road markings is presented.
1 INTRODUCTION
Road administrators require more and more objective informa
tions about their network and its surrounding environment for
various purposes : disaster management, urban planning, tourist
guidance or simply road network management are some of the
applications that demand precise city modeling and interpreta
tion. Industry also needs 3D reconstructions of large areas ; map
providers for navigation systems now include semantic data in
their bases that can be interfaced in warning or driving assistance
systems, mobile communication development needs data for ra
dio waves coverage analysis etc. These are few examples among
many fields that need augmented digital maps. Many compa
nies and research labs have then focused in the last decade on
the acquisition of mass data, developing many acquisition plat
forms. Road network surveying generally implies aerial or satel
lite multi spectral images processing but these approaches suf
fer from a lack of precision regarding road geometry, although
they provide a good classified overview of processed areas (Hat-
ger and Brenner, 2003) (Samadzadegan et al., 2009). Some re
search teams have therefore promoted fusion between terrestrial
and aerial data (Früh and Zakhor, 2004), requiring an existing
digital elevation map of the area to be processed. City modeling
is generally performed by means of vehicle borne lidar and cam
eras (Zhao and Shibasaki. 2003) (Deng et al., 2004) (Boström et
al., 2006) ; these works however do not apply on road geome
try or characterization. Some companies, cartographic institutes
and laboratories developed road dedicated vehicles, using inertial
systems and 3D lidar sensors in order to provide interpreted road
environments. StreetMapper (Barber et al., 2008) focus on eleva
tion models, ICC (Talaya et al., 2004) use stereo and (Ishikawa
et al., 2006) monocular images for automatic processes, finally
(Jaakkola et al., 2008) process lidar data as image for extracted
different kinds of road markings. (Goulette et al., 2006) only
provide automatic lidar data segmentation, performing classifica
tion of acquired scans in road, trees or obstacles. The acquisition
speed is nevertheless very low and the developed method can not
deal with rural roads, as road extraction implies curbs.
From our point of view, there were no current solution offering a
full comprehension of road environment, gathering road geome
try, road marking and traffic sign analysis in a single tool. This
is the purpose of our developments, while we focus here on road
extraction and applications from lidar data.
2 VEHICLE DESIGN AND CALIBRATION
We developed an acquisition vehicle for road surveying consist
ing in a very precise positioning system, a CCD Color camera
and 4 linear scanning sensors. A brief description of these sen
sors and the calibration methods is provided in this section.
2.1 Vehicle Specification
The positioning system consists in a Trimble Omnistar 8200-Hp
GPS receiver, combined with an Ixsea LandINS inertial measure
ment unit. This association delivers filtered 200 Hz GPS data and
can support GPS outages up to 3(KXs while presenting very small
drifts (0.005° for pitch and roll, 0.01 ° for heading, 0.7m in the xy
plane and 0.5m for the 2 coordinate). Orientation data are given
in a North East Up reference, and GPS positions are translated in
a metric coordinates system using the adequate conic projection.
As an imaging system, we use an AVT Pike F-210C, a CCD
color camera which provides Bayer filtered high definition im
ages, with a frame rate up to 30 Hz. Instead of a constant rate,
we decided to set the camera such as it takes an image every n
meters (?? is generally set to 5 m. but can be adapted depending
on environment).
Four SICK LMS-291 are installed on the roof of the vehicle (Cf.
figure 1(a)). These Laser range sensors provide 180° scans (with
0.5° angular resolution) up to 60 Hz scan rate. Their sensing
maximum range reaches 80 m with a 10 mm error and they also
can output reflectivity values (Cf. figure 6). Three of them are
looking to the ground with different orientations, the fourth one
being oriented towards the sky, in order to capture building fa
cades or trees (Cf. figure 1(b)). These sensors are controlled by
the vehicle speed, stopping the acquisition when the vehicle is
stopped.
Every data are acquired and timestamped using KI Maps software,
on a single on-board computer (Pentium IV, 2GHz) with adequate
disk space. Besides, considering the inertial navigation system
