Cuadernos Tecnolgicos de la PTC

N 09 / 2014

Automatic management of mobile LiDAR

data for road inventories
Autores:
H. Gonzlez Jorge, B. Riveiro Rodrguez, M. Varela Gonzlez, J. Martnez Snchez,
I. Puente Luna, M. Solla Carracelas, B. Conde Carnero, S. Lagela Lpez, L. Daz
Vilario, D. Roca Bernrdez, J. C. Caamao Martnez, H. Lorenzo Cimadevila, P.
Arias Snchez
Applied Geotechnology Group. School of Mining Engineering - University of Vigo.

P L ATA F O R M A T E C N O L G I C A E S PA O L A D E L A C A R R E T E R A ( P T C )

 Plataforma Tecnolgica Espaola de la Carretera (PTC). Goya, 23 - 3, 28001 Madrid.

Reservados todos los derechos. ISBN: 978-84-697-2165-0.

LA COLECCIN CUADERNOS TECNOLGICOS DE LA PTC

La Plataforma Tecnolgica Espaola de la Carretera (PTC) es el foro de encuentro apoyado
por el Ministerio de Economa y Competitividad para todos los agentes del sistema cienciatecnologa-empresa con un papel relevante en el fomento del empleo, la competitividad y el
crecimiento en el sector de las infraestructuras viarias en Espaa.
Desde su presentacin en sociedad en febrero de 2010, la PTC trabaja como una plataforma
transversal que fomenta el intercambio fluido de informacin y las discusiones a nivel
tecnolgico entre los agentes privados y pblicos del sector, con el objeto de contribuir a
que Espaa se convierta en el referente mundial en materia de tecnologas asociadas a la
carretera.
La coleccin de publicaciones Cuadernos Tecnolgicos de la PTC surge de los convenios
de colaboracin que la Plataforma mantiene con un importante nmero de instituciones
acadmicas activas en la I+D+i en materia de infraestructuras viarias. Cada Cuaderno se
incardina dentro de alguna o varias de las temticas y sub-temticas de la vigente Agenda
Estratgica de Investigacin de la Carretera en Espaa (2011-2025).

Coleccin de Cuadernos Tecnolgicos de la PTC

Ao 2013
01/2013: Tcnicas avanzadas de fusin de informacin de fuentes heterogneas para la
extraccin de informacin de movilidad en carreteras
02/2013: Software para la explotacin de datos LiDAR en carreteras
03/2013: Desarrollo de una metodologa de anlisis del coste de ciclo de vida
04/2013: Carga tarifaria y fiscal del transporte por carretera: un anlisis comparado entre
E.E.U.U. y Europa
05/2013: Captacin de energa en carretera: colectores solares asflticos
06/2013: Nuevo proceso de diseo geomtrico para unas carreteras convencionales ms
seguras
07/2013: Informe del estado del arte sobre el factor humano en la conduccin
08/2013: Optimizacin del uso de las carreteras existentes
09/2013: Diseo de estacin de carga para vehculos elctricos mediante energas renovables

Ao 2012
01/2012: Anlisis del Megatruck en Espaa
02/2012: Conceptualizacin del transporte sostenible desde el comportamiento prosocial
03/2012: Consideraciones para la modificacin de los lmites de la velocidad en base a la
accidentalidad
04/2012: Extrapolacin de materiales viarios
05/2012: Gestin de la mejora de la movilidad
06/2012: Influencia de la meteorologa adversa sobre las condiciones operacionales del trfico
y recomendaciones para la localizacin de sensores de variables atmosfricas
07/2012: Membranas flexibles ancladas al terreno para la estabilizacin de taludes en carreteras
08/2012: Priorizacin de actuaciones sobre accidentes de trfico mediante reglas de decisin
09/2012: Sistemas lidar mvil para el inventario geomtrico de carreteras

Ao 2011
01/2011: Los retos de Sistemas de adquisicin de informacin de trfico: estado actual y
futuro
02/2011: Los retos de Firmes Permeables
03/2011: Los retos del Sistema fotogramtrico para la medicin remota de estructuras en
programas de inspeccin de puentes
04/2011: Los retos de Pago por uso de las infraestructuras viarias: Estudio de los accesos a
Madrid
05/2011: Los retos del Sistema eCall: Situacin actual y estndares
06/2011: Los retos de La velocidad de operacin y su aplicacin en el anlisis de la consistencia
de carreteras para la mejora de la seguridad vial
07/2011: Los retos de Desarrollo de una metodologa de anlisis de ciclo de vida integral
especfica para carreteras
08/2011: Los retos de Control pasivo de velocidad: intervencin en tramos de acceso a
entornos urbanos

Para cualquier informacin adicional, contacte con info@ptcarretera.es o visite www.ptcarretera.es

Cuaderno Tecnolgico de la PTC N 09/2014

Automatic management of mobile LiDAR data for

road inventories
Autores:
H. Gonzlez Jorge, B. Riveiro Rodrguez, M. Varela Gonzlez, J. Martnez
Snchez, I. Puente Luna, M. Solla Carracelas, B. Conde Carnero, S. Lagela
Lpez, L. Daz Vilario, D. Roca Bernrdez, J. C. Caamao Martnez, H.
Lorenzo Cimadevila, P. Arias Snchez
Applied Geotechnology Group. School of Mining Engineering - University of
Vigo.
Agenda Estratgica
de Investigacin de la
Carretera en Espaa
(2011-2025)

Temticas:

Sub-temticas:

ITS y movilidad

Optimizacin del uso de las infraestructuras

de carreteras existentes

En colaboracin con:

Abstract
Mobile LiDAR systems are becoming increasingly popular worldwide to capture geometric
information from roads. The LiDAR systems acquire large amount of geospatial information
as points clouds, which must be processed to obtain manageable CAD or GIS product for
the end user. Traditionally data processing is done manually, however robust algorithms
are currently being developed to facilitate the extraction of information. In this paper
three algorithms are presented. The first one is a system for automatic detection of
vertical clearance in overpasses presented over roads, the second one allows the location
of the lights of a tunnel, and the third one the filtering of vehicles on the road.

Index

Index 9
1. Introduction........................................................................................................ 11
2. Automatic segmentation of road overpasses and detection of mortar efflorescence
using mobile LiDAR data........................................................................................ 13
3. Automatic detection of road tunnel luminaires using mobile LiDAR system.... 29
4. Automatic filtering of vehicles from mobile LiDAR datasets............................. 39
5. References.......................................................................................................... 49
6. Acknowledgements........................................................................................... 53

1. Introduction

Mobile LiDAR systems [1 3] have been using for several years in acquiring geospatial
data from highways and its neighborhood, both for inventory and project development
purposes. These systems allow productively acquire accurate geometric and radiometric
data from infrastructure. They have the handicap of the large amount of data generated
in a disorderly way, which must be structured to be used in engineering applications [4 8]. This manuscript presents a series of algorithms to extract information of interest from
mobile LiDAR surveys developed in roads and highways.

11

2. Automatic segmentation of road overpasses

and detection of mortar efflorescence using
mobile LiDAR data
2.1. Materials and methods
2.1.1. Area of study
Road N-120 runs between the cities of Logroo and Vigo. The road has a length of 662 km
and it has become an important passageway for other cities such as Burgos, Len, and
Ourense. During its passage throughout Ourense (Fig 1), the road depicts many overpasses
that facilitate the passage of people and goods between the two main parts of the city.
The city of Ourense experiences extreme climate, compared to the Spanish standards.
Summers with temperatures up to 40 C and winters relatively cold with minimum
temperatures around -5 C are common. Besides, the presence of Mio River running
through the city causes foggy weather in winter with high environmental moisture. These
environmental conditions require that the buildings and the structures in the city must
have special maintenance to avoid service problems and hazards.
In 1918 the New Bridge (Fig. 2) was inaugurated with the purpose of joining the two
parts of the city of Ourense to the automotive traffic. The bridge was designed by the civil
engineer Martn Daz de la Banda. It connects the streets of Curros Enrquez and Marn
Avenue and passes over the Mio River and the N-120. It shows six masonry arches and
one steel central arch with a span of 75 m. Fig. 3 shows one of the arches affected by
efflorescence, produced by the high environmental moisture and probably because of
water drainage problems in the structure.

2.2.2. Data acquisition

The mobile LiDAR system used for the survey was the Lynx Mobile Mapper from Optech
(Fig. 4) [9]. The main units of the Lynx are the navigation system Applanix Pos LV 520,
two LiDAR sensors from Optech, and four digital cameras from Jai manufacturer. All units
are assembled in a single platform to provide stability to the measurements. The data
acquisition is time stamped and the relative geometric positions among the sensors are
boresighted prior to the survey.

13

km and it has become an important passageway for other cities such as Burgos, Len,
and Ourense. During its passage throughout Ourense (Fig 1), the road depicts many
overpasses
that facilitate
the passage of people and goods between the two main
of
Cuaderno
Tecnolgico
de la PTC
N 09parts
/ 2014
the city.

Figure
RoadN120
N120in
inOurense
Ourense (red)
Fig. 1 1.Road
(red).

The city of Ourense experiences extreme climate, compared to the Spanish standards.
Summers with temperatures up to 40 C and winters relatively cold with minimum
temperatures around -5 C are common. Besides, the presence of Mio River running
through the city causes foggy weather in winter with high environmental moisture.
These environmental conditions require that the buildings and the structures in the city
must have special maintenance to avoid service problems and hazards.
In 1918 the New Bridge (Fig. 2) was inaugurated with the purpose of joining the two
parts of the city of Ourense to the automotive traffic. The bridge was designed by the
civil engineer Martn Daz de la Banda. It connects the streets of Curros Enrquez and
Marn Avenue and passes over the Mio River and the N-120. It shows six masonry
arches and one steel central arch with a span of 75 m. Fig. 3 shows one of the arches
affected by efflorescence, produced by the high environmental moisture and probably
because of water drainage problems in the structure.

Figure 2. New Bridge over N-120 road

Fig. 2. New Bridge over N-120 road.

14

Road infrastructure design for optimizing sustainability

Fig. 2. New Bridge over N-120 road.

Figure
3. Efflorescence
Efflorescence on
the masonry
masonry arches.
arches
Fig. 3.
on one
one of
of the

2.2.2. Data acquisition

The navigation system presents an accuracy of 0.020 m in XY, 0.050 m in Z, 0.005 in roll
and
andLiDAR
0.015 system
in heading
GNSS
availability).
The performance
after
60 s of
Thepitch,
mobile
used(with
for the
survey
was the Lynx
Mobile Mapper
from
GNSS
outage
0.100
m units
in XYofaccuracy,
in Z, 0.005
in roll
and pitch
Optech
(Fig. shows
4) [9]. The
main
the Lynx 0.070
are the m
navigation
system
Applanix
Pos and
0.015
heading.
LV in
520,
two LiDAR sensors from Optech, and four digital cameras from Jai

The LiDAR sensors present a maximum range of 200 m. However, due to the angle
between the LiDAR sensors and the direction of the vehicle motion (around 45), the range
of measurements regarding the navigation direction is reduced to approximately 120 m.
The range precision is 8 mm and the absolute accuracy is 5 cm. The data are digitalized
in 12 bits. The laser pulse repetition rate is programmable between 75 and 500 kHz. The
maximum pulse repetition rate was used for this work. The LiDAR sensors may measure
up to 4 echoes per pulse and then use this information for filtering vegetation and similar
applications. The scanning frequency is programmable between 80 and 200 Hz, using the
higher one for this study. The operating temperature ranges between -10 C and 40 C.
The cameras present a resolution of 5 Mpx. The speed of the electronic shutter allows up
to 3 frames per second. The field of view is 57 (horizontal) and 47 (vertical). The square
15

Cuaderno Tecnolgico de la PTC

N 09 / 2014

Figure
4. Lynx
Lynx Mobile
Mobile Mapper.
Mapper
Fig. 4.

2.2.3.
Data
processing
pixel
size
is 3.45
mm and the CCD sensor presents a 2/3 in. configuration with the Bayer
mosaic color progressive scan. The operative temperature is located between -5 C and 45
post processing
of the
Lynx
Mapper
data is doneand
after
the survey.development.
We use
C.The
Imaging
of this survey
was
notMobile
used for
data processing
algorithmic
POSPac
MMS has
5.4 to
GNSS/INS
data
during
thecloud
survey,
to createby
a each
The
data survey
an process
approximate
length
of gathered
1100 m. The
point
produced
final sensor
Smoothshows
Best Estimated
Trajectory
(SBET). To improve the accuracy of the mobile
LiDAR
around 60
million points.
GNSS, additional GNSS measurements are provided from a base station located in the
city of Ourense. The GNSS location data at the base station are recorded and timed
2.2.3.
Data
processing
during
the mission
and applied soon afterwards to the mobile GNSS data during postprocessing.
The post processing of the Lynx Mobile Mapper data is done after the survey. We use
POSPac
MMS
5.4DASH
to process
GNSS/INSalldata
to create
Following
that,
Map processes
Lynxgathered
data. Theduring
LiDARthe
rawsurvey,
measurements
area final
Smooth
Trajectory
To improve
accuracy
of the
mobile
mergedBest
withEstimated
the SBET to
calculate (SBET).
a georeferenced
pointthe
cloud,
taking into
account
theGNSS,
additional
measurements
provided
from
a base
the city of
geometricGNSS
boresight
calibration are
of the
system.
Finally,
westation
employlocated
Quick inTerrain
Ourense. The GNSS location data at the base station are recorded and timed during the
Modeler developed by Applied Imagery to visualize the processed Lynx data. Fig. 5.
mission and applied soon afterwards to the mobile GNSS data during post-processing.
shows the resultant point cloud.

Following that, DASH Map processes all Lynx data. The LiDAR raw measurements are
16

Road infrastructure design for optimizing sustainability

Figure 5. Point cloud from N-120 road

Fig. 5. Point cloud from N-120 road.

2.3. Results
andSBET
discussion
merged
with the
to calculate a georeferenced point cloud, taking into account the
geometric
of the
system.
we employ
Quick Terrain
Modeler
We have boresight
developedcalibration
a three stage
algorithm
forFinally,
extracting
the efflorescence
on road
developed by Applied Imagery to visualize the processed Lynx data. Fig. 5. shows the
overpasses. Algorithms are focused on the optimization of the point cloud, the
resultant point cloud.
automatic classification of the overpasses, and on the detection efflorescence [10]. The
flow chart is depicted in Fig. 6.

2.3.
Results and discussion
2.3.1. Optimization of the point cloud
We2.3.1.1.
have Trajectory
developed and
a three
for extracting the efflorescence on road
pointstage
cloud algorithm
normalization
overpasses. Algorithms are focused on the optimization of the point cloud, the automatic
classification
the overpasses,
and on the
detection
efflorescence
[10].
The flow
chart is
The capture of
frequency
of the GNSS/INS
system
from the
Lynx Mobile
Mapper
can be
depicted
in Fig. 6.
reconfigurable
before the survey up to a maximum of 200 Hz. Depending on the vehicle
speed, the spatial resolution of the trajectory can change. The vehicle can acquire points
even if it is stopped in a traffic jam or in traffic lights. This information does not
provide valuable data and slows down the software processing operations. To avoid the
excess data, a working speed of 80 km/h was predefined and the trajectory was
normalized to that speed. The distance between the trajectory points can be calculated
according equation d = v / 3.6f, where d is the distance in meters, v the speed of the
vehicle in km/h, and f is the acquisition frequency of the GNSS/INS in Hz. Taking into
account the aforementioned speed, a minimum distance between points of 0.11 m is
17
obtained. Therefore, trajectory points with neighbors closer than 0.11 m are removed.
Fig. 7 exhibits the results of the trajectory normalization. Since the trajectory is joined
to the point cloud through the time stamp, the corresponding point cloud was also

operation
is still high
although the data volume has been reduced forN
all09the
Cuaderno
Tecnolgico
de laenough,
PTC
/ 2014
computations approximately by 55 %.

Figure 6. Algorithm flow chart

Fig. 6. Algorithm flow chart.

2.3.1. Optimization of the point cloud

2.3.1.1. Trajectory and point cloud normalization
The capture frequency of the GNSS/INS system from the Lynx Mobile Mapper can be
reconfigurable before the survey up to a maximum of 200 Hz. Depending on the vehicle
speed, the spatial resolution of the trajectory can change. The vehicle can acquire points
even if it is stopped in a traffic jam or in traffic lights. This information does not provide
valuable data and slows down the software processing operations. To avoid the excess
data, a working speed of 80 km/h was predefined and the trajectory was normalized to that
speed. The distance between the trajectory points can be calculated according equation
d = v / 3.6f, where d is the distance in meters, v the speed of the vehicle in km/h, and f is
the acquisition frequency of the GNSS/INS in Hz. Taking into account the aforementioned
18

Road infrastructure design for optimizing sustainability

Figure 7. Input and output trajectories

Fig. 7. Input and output trajectories.

Fig. 7. Input and output trajectories.

Fig. 8. Radial filter. Input data in red and output data in blue.

Fig.8.8.Radial
Radialfilter.
filter.
Input
in and
red output
and output
data
Figure
Input
datadata
in red
data in
bluein blue.

speed, a minimum distance between points of 0.11 m is obtained. Therefore, trajectory

points with neighbors closer than 0.11 m are removed. Fig. 7 exhibits the results of the
trajectory normalization. Since the trajectory is joined to the point cloud through the time
stamp, the corresponding point cloud was also reduced and adapted to the new trajectory
data. The spatial resolution after this filtering operation is still high enough, although the
data volume has been reduced for all the computations approximately by 55 %.
19

Cuaderno Tecnolgico de la PTC

Fig. 8. Radial filter. Input data in red and output data in blue.

N 09 / 2014

Fig.9.9.Vegetation
Vegetation
filter.
Figure
filter

2.3.1.2. Radial filter

The relevant information for both road inspection and inventory is typically located in
the road vicinity and the complete range of the LiDAR data is not required for inspection
purposes, unless to perform an inspection of some elements like slopes or embankments.
In this case study, it is required to conduct the inspection of overpasses and the range was
limited to 30 m. This filtering process is also very fast, because the LiDAR data table stores
the range directly and does not require any intermediate calculation. The point cloud is
reduced from 26-788.792 points to 22.234.712 points (Fig. 8).

2.3.1.3 Vegetation filter

The area under study (Fig. 2) presents a number of trees and gardens alongside the road
which provide multiple echoes for each laser beam of the LiDAR. The vegetation filter
removes all the points that return multi-echo signal and also the neighbors inside a 5 cm
radius sphere (Fig. 9). Data were reduced up to 20.422.447 points. The filter seems to be
very useful in areas where the vegetal coverage is more representative.

2.3.2. Overpass structures segmentation and classification

Once the vegetation filter was applied to the whole dataset, the next steps consist on

20

Road infrastructure design for optimizing sustainability

Figure
Fig. 10.
10. LiDAR
LiDARangle
angleversus
versusdistance.
distance.Structures
Structuresininred
redand
andpavement
pavementiningreen
green.

the separation between the pavement and overpasses. The pavement represents a high
amount of data that must not be included in the classification process of the structures.
Fig. 10 shows the angle of LiDAR sensor L2 versus the distance run by the vehicle. The
overpasses are highlighted in red color and pavement in green. As can be observed, when
the van does not pass below an overpass, the LiDAR does not provide information about the
structures. In those situations, the laser beam is lost in the atmosphere and consequently,
there is not any laser return. Fig. 10 is used to determine what angles retrieve information
from the overpasses and separate them from the pavement information. The angle
threshold is situated between 182 and 360. It is assumed that the van is driven smoothly.
Fig. 11 shows the overpasses (in red) segmented from the pavement (in blue) with this
approach. The angles of the LiDAR selected during the boresight calibration process
cannot take clean information from the structures. The result is that some vegetation
cannot be removed from the multi-echo information of the LiDAR and some traffic signs
are also acquired as structures. A second operation of segmentation is implemented
to avoid this problem. The segmented data correspond only to the central part of the
structure,
vertical
clearance
11 green
dataset).
segmentation
is
Fig. 11. typically
Structuresthe
(red)
segmented
from (Fig.
the pavement
(blue)
using Such
the LiDAR
angle
evaluated
using
a vertical
angular
from the
LiDARcolored.
sensors of 6.
approach.
Vertical
clearance
of theslice
structures
is green
On the other hand, classification is a time consuming computational process for large21
datasets. In this case, the classification time is considerably reduced using the dataset
that represents the time stamp for the previously segmented data using the vertical
LiDAR slice of 6. Fig. 12 represents a histogram of the dataset. After that, a K-means

Cuaderno Tecnolgico de la PTC

N 09 / 2014
Fig. 10. LiDAR angle versus distance. Structures in red and pavement in green.

Figure
segmented
from
thethe
pavement
(blue)
using
the the
LiDAR
angleangle
approach.
Fig. 11.
11.Structures
Structures(red)
(red)
segmented
from
pavement
(blue)
using
LiDAR
Vertical
clearance
of theofstructures
is green
colored
approach.
Vertical
clearance
the structures
is green
colored.

On the other hand, classification is a time consuming computational process for large
datasets. In this case, the classification time is considerably reduced using the dataset
that represents the time stamp for the previously segmented data using the vertical
LiDAR slice of 6. Fig. 12 represents a histogram of the dataset. After that, a K-means
clustering algorithm is applied for classification.

Figure
the clearance
clearance of
the structures.
structures
Fig. 12.
12. Histrogram
Histrogramof
of the
the time
time stamp
stamp of
of the
of the

The unsupervised classification is a highly computer-automated procedure. This method

Onautomatically
the other hand,
classification
a timeofconsuming
process tofora large
divides
the range ofis values
the image computational
into clusters according
datasets.
In this
case, the
classification
time is considerably
reduced
dataset that
statistical
procedure,
without
any predefined
knowledge. The
idea using
of thethe
K-means
represents
stamp
for the previously
segmented
data
usingown
thedata,
vertical
algorithm the
is totime
provide
a classification
of information
according
to their
basedLiDAR
slice
6. Fig.and
12 represents
histogram
ofnumerical
the dataset.
After
that,the
a K-means
on of
analyses
comparisonsa between
their
values.
Thus,
algorithmclustering
will
algorithm is applied for classification.
provide an automatic sorting without pre-existing classification. Because of this feature,
the K-means is regarded as an unsupervised data mining algorithm.
22

In the K-means algorithm, the user can specify various parameters, including the desired
number of clusters and maximum number of iterations allowed in the program. The
main idea is to define K centroids, one for each class. The next step is to take each time

Road infrastructure design for optimizing sustainability

Figure
of of
thethe
structures
(from(from
one to
four)
the K-means
algorithm
Fig. 13.
13.Classification
Classification
structures
one
to according
four) according
the K-means
and and
timetime
stamp
data.
UTMUTM
coordinates
are used
algorithm
stamp
data.
coordinates
are used.

Figure 14. Example of the point cloud of one of the bridges automatically extracted and
Fig. 14. Example of the point cloud of one of the bridges automatically extracted and
classified from the datasheet
classified from the datasheet.
23

2.3.3. Efflorescence detection

Cuaderno Tecnolgico de la PTC

N 09 / 2014

The unsupervised classification is a highly computer-automated procedure. This method

automatically divides the range of values of the image into clusters according to a statistical
procedure, without any predefined knowledge. The idea of the K-means algorithm is to
provide a classification of information according to their own data, based on analyses
and comparisons between their numerical values. Thus, the algorithm will provide an
automatic sorting without pre-existing classification. Because of this feature, the K-means
is regarded as an unsupervised data mining algorithm.
In the K-means algorithm, the user can specify various parameters, including the desired
number of clusters and maximum number of iterations allowed in the program. The
main idea is to define K centroids, one for each class. The next step is to take each time
stamp data belonging to a given data set and associate it with the nearest centroid, and
at this point, the centroid should be recalculated. Then, we have K new centroids and a
new connection has to be made between the same set of data points and the new near
centroid. These steps define a new iteration. As this is a iterative process, we can see that
the K centroids change their location step by step until no further change is possible.
Taking into account this data, the maximum, minimum and average of each class are easily
obtained. Once the maximum and minimum time stamp is defined for each cluster, these
values are used in the original dataset of the point cloud, with the pavement preciously
removed, to obtain the complete structure. It should be noted that the information of the
two LiDAR sensors must be used simultaneously to obtain a good coverage of the overpass
structure. Results are shown in Figs. 13 and 14. The classification is 100 % successful, because
the structures are enough isolated according to the time stamp value.

2.3.3. Efflorescence detection

The LiDAR data represent a 12 bit intensity pattern, although the main part of the data is
located in their 255 first values (similar to 12 bit). Fig. 15 shows a histogram of the data. It is
known from the previous works that efflorescence and moisture provide lower reflection
than concrete and masonry stone. Thus, the solution for the automatic detection of
efflorescence is implemented taking into account an intensity threshold. It is very important
to execute the previous extraction of the pavement data (Fig. 10 green highlighted),
because the intensity response of the pavement is similar to the efflorescence and can
provide false positives.
Fig. 16 shows a profile of LiDAR data at a height of 6.5 m (red profile) in one of the
overpasses with efflorescence. The graph shows the distance run by the vehicle and the
corresponding intensity recorded by the LiDAR. Intensity is smoothed using a moving
average filter of five elements. The figure also includes an image of the point cloud colored
using intensity values. The data depicts that efflorescence provides lower intensity values

24

detection of efflorescence is implemented taking into account an intensity threshold. It

is very important to execute the previous extraction of the pavement data (Fig. 10
green highlighted), because the intensity response of the pavement is similar to the
Road
infrastructure
design
for optimizing
sustainability
efflorescence
and
can provide
false positives.

Figure
15. Intensity
Intensity histogram
histogram of
of one
one of
ofthe
thestructures.
structures
Fig. 15.

Fig. 16 shows a profile of LiDAR data at a height of 6.5 m (red profile) in one of the
overpasses with efflorescence. The graph shows the distance run by the vehicle and the
corresponding intensity recorded by the LiDAR. Intensity is smoothed using a moving
average filter of five elements. The figure also includes an image of the point cloud
colored using intensity values. The data depicts that efflorescence provides lower
intensity values (blue highlighted). Taking into account this data a threshold value was
selected (green line in the figure). Fig. 17 shows the efflorescence points extracted from
the point cloud using the intensity threshold (red colored). The data are in agreement
with the real situation of the bridge and the interpretation done by the experts.

Figure 16. Intensity profile of the structure at 6.5 m height (red). Efflorescence in the point cloud
Fig. 16. Intensity profile of the structure at 6.5 m height (red). Efflorescence in the point
in blue

cloud in blue.

25

Fig. 16.Tecnolgico
Intensity profile
of the structure at 6.5 m height (red). Efflorescence inN
the09
point
Cuaderno
de la PTC
/ 2014
cloud in blue.

Figure
Fig. 17.
17. Efflorescence
Efflorescencedetection
detectionininred
red.

(blue highlighted). Taking into account this data a threshold value was selected (green line
in the figure). Fig. 17 shows the efflorescence points extracted from the point cloud using
the intensity threshold (red colored). The data are in agreement with the real situation of
the bridge and the interpretation done by the experts.

2.4. Conclusion
A set of algorithms for the automatic classification and detection of efflorescence using
mobile LiDAR data is depicted. The mobile LiDAR datasets are first automatically normalized
to avoid the excessive data when the vehicle is stopped or traveling slow. The point cloud
is reduced by 55 %. Then, a radial filter is applied to remove all points with a range larger
than 30 m, assuming that all the information of interest is within this range. The point
cloud is reduced up to 22 million points. Most of the vegetation data were removed using
a filter based in the second echo of the LiDAR data. The achieved reduction is about 1.9
million points, up to 20 million points.
Two different approaches are studied for the segmentation of the structures, using the
information of the LiDAR angle. The first one uses the limiting angle of the pavement
and assumes that the points that are not inside this limit correspond to overpasses.
26

Road infrastructure design for optimizing sustainability

This algorithm segments properly the pavement, although does not provide a clean
point cloud with the overpasses. Many points from non-filtered vegetation and poles
are taken as structures. The second approach limits the angle to find a structure to 6 in
the perpendicular of the vehicle. The fine slice of data avoids completely trees or poles.
The complete structure can be then regenerated taking into account the minimum and
maximum time stamp of the data for each cluster (single overpass) and all the point cloud
data inside this range. The pavement is previously removed from the point cloud. A Kmeans classifier is used in this step.
Finally, the detection of the efflorescence is performed using the intensity data from the
LiDAR. The efflorescence diminishes the reflected radiation from the LiDAR and can be
easily segmented from the surrounding point cloud.

27

3. Automatic detection of road tunnel

luminaires using mobile LiDAR system.
3.1. Materials and methods
3.1.1. Test site and data collection
The tunnel selected in this study is the largest engineering project and one of the
cornerstones of the current highway A-52, that connects the principal cities in southern
Galicia, Spain. The O Folgoso tunnels course runs in the municipality of A Caiza, with a
total length of 2500 m (Fig. 18). The tunnel is monitored by a control center that checks the
ventilation, light, signaling, communications, and fire detection systems among others.
The route (Fig. 18) was driven at a theoretical constant speed of 66 km/h and at a specific
time in the day with low traffic density and no obstructions in order to ensure maximum
coverage and minimizing shadow areas.

3.1.2. Data post processing

The GPS and IMU data was post processed using Applanixs POSPAC MMS 5.4 which
processes the GPS and IMU data into a trajectory. There was no GPS collection inside
the tunnel and consequently, the system navigates using the inertial navigation system
(INS) and the odometer (DMI). In spite of the quality of the inertial navigation, error for
absolute positioning increase up to 0.47 m (Fig 19). Although this accuracy seems enough
for a light inventory, results could be improved if necessary, using geo-referenced control
points every 25 m in the data processing phase as long as the data has good initialization,
good finalization and good inertial and DMI measurements during tunnel data acquisition.
The result of POSPAC was used to compute the positioning of the point cloud processed
using Optechs DASHMAP for converting the raw laser data to .las or .xyz files. Also note
that LiDAR intensity is recorded and this data will be used to color those areas of the point
cloud without RGB information. Thanks to the camera boresighting, that ensures that the
cameras are correctly calibrated to the IMU and the LiDAR sensors, it is possible to obtain
the texture mapping of each laser point.

29

Galicia, Spain. The O Folgoso tunnels course runs in the municipality of A Caiza,
with a total length of 2500 m (Fig. 18). The tunnel is monitored by a control center that
Cuaderno
Tecnolgico
de la PTClight, signaling, communications, and fire detectionNsystems
09 / 2014
checks
the ventilation,
among others.

Figure
18. 18.
O Folgoso
tunnel:tunnel:
(a) inside
tunnelthe
looking
(source:Vigo
Anonymous)
Fig.
O Folgoso
(a)the
inside
tunneltowards
lookingVigo
towards
(source:
and (b) top
of the
point
cloud
acquired
with the with
Lynx the
mobile
Anonymous)
andview
(b) top
view
of the
point
cloud acquired
LynxLiDAR
mobile LiDAR.

The route (Fig. 18) was driven at a theoretical constant speed of 66 km/h and at a
specific time in the day with low traffic density and no obstructions in order to ensure
maximum coverage and minimizing shadow areas.
3.1.2. Data post processing
The GPS and IMU data was post processed using Applanixs POSPAC MMS 5.4 which
processes the GPS and IMU data into a trajectory. There was no GPS collection inside
the tunnel and consequently, the system navigates using the inertial navigation system
(INS) and the odometer (DMI). In spite of the quality of the inertial navigation, error for
absolute positioning increase up to 0.47 m (Fig 19). Although this accuracy seems
enough for a light inventory, results could be improved if necessary, using georeferenced control points every 25 m in the data processing phase as long as the data has
good initialization, good finalization and good inertial and DMI measurements during
tunnel data acquisition.

Figure
and
number
of of
GPS
satellites
during
the data
collection
Fig. 19.
19.Smoother
Smootherperformance
performancemetrics
metrics
and
number
GPS
satellites
during
the data
collection.

The result of POSPAC was used to compute the positioning of the point cloud
processed
using Optechs DASHMAP for converting the raw laser data to .las or .xyz
30
files. Also note that LiDAR intensity is recorded and this data will be used to color
those areas of the point cloud without RGB information. Thanks to the camera
boresighting, that ensures that the cameras are correctly calibrated to the IMU and the

Road infrastructure design for optimizing sustainability

3.2. Description of the algorithm

It is well known that most of the features present in infrastructures, such as lanes, barriers,
traffic poles, or emergency exits are important for road inventory purposes. Each feature
is described with a number of distinguishing attributes and its relation with other features.
To detect the lighting system in the tunnel effectively, some attribute constraints should
be considered first to avoid false positives [11].
Those attributes include position and the color for the algorithm explained below. The
luminaires are expected to be at higher locations at the side of the tunnel. By incorporating
this information, the luminaires can be distinguished from for example, vehicle lights or
vehicle shadows. Moreover, the color pattern of objects can be used to identify a large
number of features. Inside the tunnel, the luminaires have fixed color patterns, brighter
than its surroundings, which can be used to identify them. However, we should be aware
that this algorithm will fail if the luminary is switched off.
For that reason, and in order to guarantee the success of this procedure, the experts
must assure that all the luminaires are working during the first data acquisition. Under this
premise we could detect in the following inspections if a luminaire is still working, or on
the contrary, if it is turned off or damaged. Being able to detect if a luminaire is damaged
or not working will definitely improve the productivity of tunnel inspections.
The algorithm is given as follows (Fig. 20). The input is a colored point cloud acquired
by laser scanning inside a tunnel. The geometric and radiometric characteristics of the
aforementioned point cloud are used to filter out the luminaires from the total amount of
scanner data. The first step (stage 1) simply consists of the point cloud pre-segmentation
based on height values. The original axis can be extracted from the vehicle trajectory
(obtained by the inertial navigation system located at the top of the vehicle). Then, all the
points from the point cloud that are at least 2 m above the ground level are considered for
the new stage (stage 2). This user-defined value is specifically determined for each case
study, by measuring the height of the first luminaires manually. Subsequently, points are
uniformly sampled and those which verify the conditions set by the chosen color threshold
are considered as tunnel luminaires. Thresholds for classification are extracted from the
color histrogram analysis taken of 8 images for each component (red, green, and blue)
gives the darkness color (no light, considered as black), while the RGB value of 255, 255,
255 means pure white.
The third stage of the algorithm enhances the raw solution obtained after stage 2.
The rapid movement of the vehicle and a slow shutter speed result in blurred images.
Accordingly, the luminaire dimensions are distorted and a post-refined of their lengths is
necessary. The correction uses known values of vehicle speed and camera shutter speed
to automatically compute the elongation of the luminaire and recalculate its original size.
31

Cuaderno Tecnolgico de la PTC

N 09 / 2014

Figure 20. Algorithm workflow applied to the tunnel data

Fig. 20. Algorithm workflow applied to the tunnel data.

3.3. Results and discussion
32

The strip from O Folgoso tunnel used for this test is about 0.5 km long, containing
10.0230895 laser points in total. Processing of single part from the tunnel data is

and ceiling are colored by LiDAR intensity instead.

Then the points are classified into luminaires if the RGB color value is above the
threshold
defined
by the
user. Previously,
only points with a certain height above the
Road
infrastructure
design
for optimizing
sustainability
ground level (2 m) were considered (Fig. 22) as described next.

Fig. 21.
21.O O
Folgoso
tunnel:
3D colored
tunnel
point
with highlighted
luminairesby a
Figure
Folgoso
tunnel:
(a) 3D(a)
colored
tunnel point
cloud
with cloud
luminaires
highlighted
by a redand
rectangle
(b) luminaires
two tunnel recorded
luminaires
recorded
with system
the camera
red rectangle
(b) twoand
tunnel
with
the camera
from the
systemLynx
from the Lynx.

Although the vehicle travels at constant speed, authors assume that the cars acceleration
is changing slightly at certain moments. Therefore, each luminaire will have a specific
elongation based on its instantaneous vehicle speed, which can be computed thanks to
the inertial system. On the other hand, the camera shutter speed was manually controlled
during the data acquisition and its value is kept constant at 1/15 s for these tests.
A final stage 4 is included in the algorithm. Detailed tunnel luminaire inventories can
be created and saved using a GIS database. Within this technology, the luminaires are
normally defined by a single point that will be the average of the input points. We easily
calculate the centroid for each luminaire as the mean position of all the points belonging
to the previously segmented and refined luminaire. Thanks to it, several metadata which
captures relevant information for luminaires can be linked in an efficient way to their
Fig. 22.coordinates.
Stage 1. Automatic height filtering tool: (a) before and (b) after being applied
absolute
to the tunnel point cloud.
The color of the tunnel is visually divided into two groups: the luminaires, with white
3.3.
Results and discussion

color, and the surrounding environment, with dark color. The distribution of the digital
level
of the
pixels
recordedtunnel
with theused
camera
deployed
on the
is shown
in
The
strip
from
O Folgoso
for system
this test
is about
0.5MLS
km long,
containing
Fig. 23. laser points in total. Processing of single part from the tunnel data is illustrated
10.0230895
in Fig. 21. The laser points were colored at first. Note that there is no RGB information
about the ground level and tunnel ceiling due to the camera specifications and camera
system configuration on the vehicle, therefore points representing ground and ceiling are
colored by LiDAR intensity instead.

Then the points are classified into luminaires if the RGB color value is above the threshold
33

Fig. 21. O Folgoso tunnel: (a) 3D colored tunnel point cloud with luminaires
highlighted
by a red
Cuaderno
Tecnolgico
de rectangle
la PTC and (b) two tunnel luminaires recorded with the camera
N 09 / 2014
system from the Lynx.

Fig. 22.
22. Stage
Stage 1.
1.Automatic
Automaticheight
heightfiltering
filteringtool:
tool:(a)(a)
before
and
after
being
applied
Figure
before
and
(b)(b)
after
being
applied
to the
to the tunnel point cloud.
cloud
The color of the tunnel is visually divided into two groups: the luminaires, with white
color, and the surrounding environment, with dark color. The distribution of the digital
level of the pixels recorded with the camera system deployed on the MLS is shown in
Fig. 23.

FigureFig.
23.23.
Stage
2. Histrogram
withwith
thethe
digital
level
information
contained
in the
tested
data
Stage
2. Histrogram
digital
level
information
contained
in the
tested
data.

defined by the user. Previously, only points with a certain height above the ground level (2
The black and white color distribution results in two peaks in the histrogram due to its
m) were considered (Fig. 22) as described next.
different properties. This case can be considered as a mixture of two different uni-modal
distributions, each one representing a different skewed distribution: positively (left side)
The color of the tunnel is visually divided into two groups: the luminaires, with white color,
and negatively (right side). The two peaks of the bimodal histrogram are approximately
and the
surrounding environment, with dark color. The distribution of the digital level of
at
7.7
and 255. Considering
that thesystem
difference
betweenon
thethe
two
modes
is big enough,
the pixels recorded
with the camera
deployed
MLS
is shown
in Fig. 23.
the threshold for luminary detection can be easily set on a certain value within the range
150 220, and the result will not vary significantly.
34

Fig. 24 and 25 illustrate the lightning system recognition of a tunnel part from the data
set. In Fig. 26, the results of our extraction method are checked by back projecting the
detected luminaires into the original point cloud. Stages 3 and 4 are finally executed for

Road infrastructure design for optimizing sustainability

Figure
24. Segmented
Segmented tunnel
tunnel luminaires
luminaires after
after stages
stages 11 and
and 2.
2
Fig. 24.

The black and white color distribution results in two peaks in the histrogram due to its
different properties. This case can be considered as a mixture of two different uni-modal
distributions, each one representing a different skewed distribution: positively (left side)
and negatively (right side). The two peaks of the bimodal histrogram are approximately
at 7.7 and 255. Considering that the difference between the two modes is big enough, the
threshold for luminary detection can be easily set on a certain value within the range 150
220, and the result will not vary significantly.
Fig. 24 and 25 illustrate the lightning system recognition of a tunnel part from the data set.
In Fig. 26, the results of our extraction method are checked by back projecting the detected
luminaires into the original point cloud. Stages 3 and 4 are finally executed for refining the
existing segmentation of the luminaires and calculating their centroids afterwards. Note
that the real shape of the luminary is squared although it seems to be rectangular in the
photos taken. As a consequence, this pattern is repeated in the colored point cloud that
35

Fig. 25. Randomly tunnel luminaire in detail.

Cuaderno Tecnolgico de la PTC

N 09 / 2014

Fig. 24. Segmented tunnel luminaires after stages 1 and 2.

Fig. 25.
Figure
25. Randomly
Randomly tunnel
tunnel luminaire
luminaire in
in detail.
detail

Figure
Fig. 26.
26. Refined
Refined segmentation
segmentation and
and centroid
centroid calculation
calculationofoftunnel
tunnelluminaires
luminairesafter
afterstages
stages
3 and 4
3 and 4.

3.3. Conclusions
36

Mobile Laser Scanning technology collects immense amounts of accurate information

in the field, reducing survey time and minimizing the potential safety risks to ground

Road infrastructure design for optimizing sustainability

we used as an input. This problem is corrected based on the known values of vehicle
speed and camera shutter speed. The algorithm automatically substracts the motion blur
length error to compute the real size of the luminary. Authors have tested 62 luminaires,
obtaining a mean value of 0.378 m length (0.380 m is the nominal value).

3.3. Conclusions
Mobile Laser Scanning technology collects immense amounts of accurate information in
the field, reducing survey time and minimizing the potential safety risks to ground crews
traditionally collecting this type of data. In this paper we presented an algorithm for the
automatic detection of tunnel luminaires using point clouds obtained by MLS.
Preliminary results from the segmentation have been good. More importantly, this
straightforward algorithm is extensible as long as the luminaires are turned on and some
geometric constrains are assumed. For example, the algorithm workflow presented here
could be applied successfully to recognize street lamps in urban areas only when the
surveying is performed during the night. Otherwise, the stage 2 would fail as the differences
in appearance in color remain ambiguous. Also, by assuring that all the luminaires are
turned on during the first data acquisition, it will be possible later to detect and locate
where a luminaire is turned off or just not working properly.
This simple and practical method requires users to adjust only two thresholds (height and
color intensity). It detected the lighting system with 100 % accuracy in a total time of 95 s,
exhibiting a good processing time for large datasets.

37

4. Automatic filtering of vehicles from

mobile LiDAR datasets
4.1. Test site
The road strip selected for this research corresponds to the N-120 passing through the
city of Ourense in Spain (Fig. 27). The length of the strip used in this work is 1.5 km. This
road has high concentration of traffic and allows the testing of the algorithms for vehicle
filtering in demanding conditions. In addition, it reproduces a conventional European
road, which gives reproducibility to the study.

4.2. Algorithm description

The algorithm is divided in two main steps. One consists of the normalization and
segmentation of the road and another in the detection of the contour of the vehicle
and removing the corresponding LiDAR echoes. The input data for the algorithm are the
trajectory and LiDAR datasets and the output the point cloud without the echoes from the
vehicles. The road surface is finally regenerated using commercial software [12]. Flowchart
with the description of the algorithm is shown in Fig. 28.

4.2.1. Road normalization and segmentation

The trajectory was normalized to a speed of 80 km/h and a spatial resolution between
points close to 10 cm is obtained. Since trajectory data is linked to the point cloud through
the time stamp, the normalization of the trajectory can be transferred to the point cloud.
The dataset is reduced applying a radial filter in the neighborhood of the road (distance of
20 m). The point cloud is reduced from 114.9 LiDAR points up to 49.6 LiDAR points.

4.2.2. Location of the contour of the vehicle and cleaning

The location of the contour of the vehicle is done using a moving slope algorithm. The
algorithm uses least squares linear fitting calculation that works in a mobile form with
-groups of 5 points on each profile extracted from the point cloud. The point cloud is
formed by profiles of points coming from the rotation of the LiDAR mirror and the filtering
operations previously done. Because the LiDAR point clouds store range and angle
39

Cuaderno Tecnolgico de la PTC

Fig. 27. Road strip. Image (top) and point cloud (bottom). Vehicles in red.

Figure 27. Road strip. Image (top) and point cloud (bottom). Vehicles in red

40

N 09 / 2014

Road infrastructure design for optimizing sustainability

Figure 28. Flowchart summarizing the algorithmic

Fig. 28. Flowchart summarizing the algorithmic.

information in the dataset, it is very easy to automatically segment each of the profiles
and
runstrategy
the algorithm
a largeisslope,
denoting
of the
vehicle,
The
of the until
algorithm
shown
in Fig.the
29.beginning
Three main
areas
can isbefound.
Dataset
is organized
in 7 columns:
timevehicle
stamp,detection
x coordinates,
y coordinates,
z coordinates,
distinguished.
The threshold
for the
was fixed
to 75 based
on the
intensity,
angle.
expertiserange,
of theand
authors,
because the majority of the vehicles tend to show 90 in the
side vehicle angle. Moreover, there are not elements in the road that could present
The
strategy
the 75
algorithm
is shown
29. Three
main
areas
can be
angles
lowerofthan
that could
appearinasFig.
a false
positive
of the
vehicle.
Thedistinguished.
angular
The threshold for the vehicle detection was fixed to 75 based on the expertise of the
values are calculated automatically from the obtained slope. The mathematical value of
authors, because the majority of the vehicles tend to show 90 in the side vehicle angle.
the slope comes
the elements
linear fitting
theroad
closest
5 points
the profile.
Moreover,
there from
are not
in of
the
that
could of
present
angles lower than 75
that could appear as a false positive of the vehicle. The angular values are calculated
The vehicle limit was calculated by the intersection of two lines corresponding to
automatically from the obtained slope. The mathematical value of the slope comes from
pavement areas. Fig. 30 shows the results of the algorithm behavior for a set
thevehicle
linearand
fitting
of the closest 5 points of the profile.
of data that involve 2 vehicles. Once the limit of the vehicle is detected the points of the
profile with a Z value higher than the Z coordinate of the limit are removed. Fig. 31
shows the final result after the removal. The application of the algorithm is limited to41
the area inside the road lines, which can be easily detected using their high reflectance.
The location of the contour of the vehicle and cleaning runs unsupervised without

Cuaderno Tecnolgico de la PTC

Figure
29. Strategy
Strategy of
of the
the slope
slopemoving
movingalgorithm.
algorithm
Fig. 29.
Fig. 29. Strategy of the slope moving algorithm.

Figure 30. Location of the contour of the vehicle (red)

Fig. 30. Location of the contour of the vehicle (red).
Fig. 30. Location of the contour of the vehicle (red).
42

N 09 / 2014

Road infrastructure design for optimizing sustainability

FigureFig.
31. Example
withwith
the data
removed
from
the the
vehicles
(red)
31. Example
the data
removed
from
vehicles
(red).

The vehicle limit was calculated by the intersection of two lines corresponding to vehicle
Once the vehicle
is cleaned,
the user
can regenerate
the point cloud
usingfor
surface
and pavement
areas. Fig.
30 shows
the results
of the algorithm
behavior
a setfitting
of data
algorithms.
These
algorithms
are
simple,
have
been
used
for
many
years,
and
are
that involve 2 vehicles. Once the limit of the vehicle is detected the points of the profile
with aintegrated
Z value higher
than
the Z coordinate
the32).
limit are removed. Fig. 31 shows the final
in much
commercial
softwareof
(Fig.
result after the removal. The application of the algorithm is limited to the area inside the
4.3. Results
road lines,
which and
can discussion
be easily detected using their high reflectance. The location of the
contour of the vehicle and cleaning runs unsupervised without human intervention.
To check the behavior of the filter three tests were done. The first one consisted of
of isthe
false positives
by the method;
second
focuses
onfitting
the
Once detection
the vehicle
cleaned,
the userprovided
can regenerate
the pointthe
cloud
using
surface
evaluation
ofalgorithms
a longitudinal
profile have
of thebeen
regenerated
third test
algorithms.
These
are simple,
used forpoint
manycloud,
years,and
and the
are integrated
evaluated
the cross
sections.
in much
commercial
software
(Fig. 32).

4.3. Results and discussion

To check the behavior of the filter three tests were done. The first one consisted of detection
of the false positives provided by the method; the second focuses on the evaluation of a
longitudinal profile of the regenerated point cloud, and the third test evaluated the cross
sections.
The road strip used for the test includes 43 vehicles inside the traffic lines and 6 vehicles
parked in a service road disposed parallel to the main road. The whole points cloud was
reviewed by the authors and no false positives appeared inside the traffic lines. This is due
to the robustness of the algorithm which limits the searching to only inside the limits of
43

Cuaderno Tecnolgico de la PTC

N 09 / 2014

Figure 32. Example of the pavement holes regenerated using commercial software. Top left
Fig. 32.
Example
pavement
software.
image
showofthethe
points
aroundholes
eachregenerated
hole used tousing
fit thecommercial
surface. Top-right
and Top
bottom
left image
show
thethe
points
eachthehole
used to fitVehicles
the surface.
andnext
images
show
new around
points from
regeneration.
on theTop-right
service road,
bottomtoimages
show
theare
newnot
points
from(white)
the regeneration. Vehicles on the service road,
the main
road
detected
next to the main road are not detected (white).

the road line. Due to the research strategy of the algorithm, vehicles outside the road lines
are not detected (Fig. 32). This is a limitation difficult to solve in a robust manner.
The road strip used for the test includes 43 vehicles inside the traffic lines and 6
Fig. 33 and 34 show two longitudinal profiles done in a part to compare the reconstructed
vehicles parked in a service road disposed parallel to the main road. The whole points
area with an area that does not need reconstruction. As can be observed the reconstructed
clouddoes
was not
reviewed
the authors
and no false
appeared
inside
traffic
profile
show by
appreciable
differences
withpositives
the pavement
data.
Thethe
reconstructed
lines.
is dueimportant
to the robustness
of to
thethe
algorithm
theThe
searching
to only
holes
doThis
not show
noise due
filteringwhich
of thelimits
surface.
non-reconstructed
insideshows
the limits
of the
road line.
Due to
the research
strategy
of the
algorithm,
profile
higher
density
of points
because
it is closer
to the
mobile
LiDARvehicles
system. The
outside
theprofiles
road lines
are not
detected
(Fig. 32).
Thisroad.
is a limitation difficult to solve in
slope
in the
comes
from
the camber
of the
a robust manner.
44 33 and 34 show two longitudinal profiles done in a part to compare the
Fig.
reconstructed area with an area that does not need reconstruction. As can be observed
the reconstructed profile does not show appreciable differences with the pavement data.

Road
design for
optimizing
Theinfrastructure
non-reconstructed
profile
showssustainability
higher density of points because it is closer to the
The
non-reconstructed
profile
shows
higher
density
of from
pointsthe
because
is the
closer
to the
mobile LiDAR system. The slope in the
profiles
comes
camberit of
road.
mobile LiDAR system. The slope in the profiles comes from the camber of the road.

Fig. 33.
33. Longitudinal
Longitudinalsections
sectionstotocompare
compare
reconstructed
(blue)
non-reconstructed
Figure
reconstructed
(blue)
andand
non-reconstructed
Fig.
33.
Longitudinal
sections
to
compare
reconstructed
(blue)
and
non-reconstructed
pavement
(green)(green)
pavement
pavement (green)

Fig. 34.
34. Reconstructed
Reconstructedprofile
profile(blue)
(blue)and
andnon-reconstructed
non-reconstructedprofile
profile(green)
(green).
Figure
Fig. 34. Reconstructed profile (blue) and non-reconstructed profile (green).

Cross sections of the road were automatically generated perpendicular to the trajectory
of the mobile LiDAR system. Distance between each section is 30 cm. The cross sections
consist of slices from the point cloud with a thickness of 10 cm. Cross sections algorithm
selects the points whose distance to a plane perpendicular to the trajectory is lower than
5 cm. The distance between the generated planes is 30 cm.
Fig. 35 shows an example of three different parts of the road and the associated cross
sections. The first one (Fig. 35 A) represents the cross section of a hole and the second
45

Cuaderno Tecnolgico de la PTC

N 09 / 2014

Fig. 35.
35. Cross
Crosssections.
sections.(A)
(A)Pavement
Pavementhole,
hole,
pavement
hole
with
vehicles,
(C)
Figure
(B)(B)
pavement
hole
with
vehicles,
andand
(C) area
area without
withoutholes
holesororvehicles.
vehicles
Cross sections of the road were automatically generated perpendicular to the trajectory
one (Fig. 35 B) the beginning of a vehicle. It can be observed vertical points (green) from
of the mobile LiDAR system. Distance between each section is 30 cm. The cross
the echoes of the LiDAR from the vehicle. The reconstructed profile (red) shows a robust
sectionsand
consist
of slices from
the point
cloud with
a thickness of 10 cm. Cross sections
behavior
regenerates
the road
in a correct
manner.
algorithm selects the points whose distance to a plane perpendicular to the trajectory is
lower
than 5 area
cm. The
between
the generated
is 30 cm.
Each
vehicle
candistance
be divided
always
in two planes
parts when
a cross section is done
perpendicular to it. One of the parts presents the hole and the other one also includes
Fig. 35from
shows
example
of three
different
parts
of the
and the associated
crosssensor
echoes
theanvehicle.
This
situation
comes
from
theroad
orientation
of the LiDAR
sections.
The first
one (Fig.
35 regardless
A) represents
crossdirection)
section ofduring
a hole and
second
heads
(forming
an angle
of 45
of the
vehicle
datathe
acquisition.
one (Fig. 35 B) the beginning of a vehicle. It can be observed vertical points (green)
from
echoestwo
of the
LiDAR
from without
the vehicle.
Theor
reconstructed
profile
showsare
a very
Fig.
35Cthe
shows
cross
sections
holes
vehicles. The
two(red)
profiles
similar.
robust behavior and regenerates the road in a correct manner.
Each vehicle area can be divided always in two parts when a cross section is done
perpendicular to it. One of the parts presents the hole and the other one also includes
echoes from the vehicle. This situation comes from the orientation of the LiDAR sensor
46 (forming an angle of 45 regardless of vehicle direction) during data acquisition.
heads
Fig. 35C shows two cross sections without holes or vehicles. The two profiles are very

Road infrastructure design for optimizing sustainability

4.4. Conclusions
An algorithm is developed to filter echoes from vehicles in mobile LiDAR datasets. First
step of the algorithm includes road normalization and segmentation. Second step includes
the detection and cleaning of the vehicles. Detection is done using a least squared fitting
calculation that works in mobile form with groups of 5 points on each point cloud profile.
The algorithm assumes that all the vehicles are inside the reflective lines of the road.
Once the vehicle is detected and removed, the remaining holes can be regenerated using
surface fitting algorithms with the neighborhood points.
The algorithms were tested for 43 vehicles inside the traffic lines and 6 vehicles outside
in a service road. All the vehicles inside the traffic lines were successfully filtered. The
regenerated pavement is compared with the situations where no regenerations is required
by means of longitudinal profiles and cross sections. In both situations the point cloud of
the regenerated pavement is in agreement with the parts of the road where regeneration
was not required.

47

5. References

[1] G. Petri, Mobile Mapping systems: An introduction to the technology. Geoinformatics. 13(1)
(2010) 32 43.
[2] I. Puente, H. Gonzlez-Jorge, J. Martnez-Snchez, and P. Arias, Review of mobile mapping
and surveying technologies, Measurement 46(7) (2013) 2127- 2145.
[3] L. Graham, Mapping systems overview, Photogrammetric Engineering and Remote Sensing
76(3) (2010) 222 228.
[4] M. Varela-Gonzlez, H. Gonzlez-Jorge, B. Riveiro, and P. Arias, Performance testing of LiDAR
exploitation software, Computers & Geoscience 54 (2013) 122 -129.
[5] A. Jaakkola, J. Hyyppa, and A. Kukko. Retrieval algorithms for road surface modeling using
laser-based mobile mapping, Sensors. 8(9) (2008) 5238 5249.
[6] J. Herandez and B. Marcotegui. Filtering artifacts and pavement segmentation from mobile
LiDAR data, in: Laser Scanning IAPRS2009, 2009, XXXVIII Part 3/W8, pp. 329
333.
[7] C. Mc Elhinney, P. Kumar, C. Cahalane, and T. Mc Carthy, Initial results from the European road
safety inspections (EURSI) mobile mapping project, International Archives of Photogrammetry,
Remote Sensing and Spatial Information Sciences. XXXVII Part 5 (2010) 440 445.
[8] A. Boyko and T. Funkhouser, Extracting roads from dense point clouds in large scale urban
environment, ISPRS Journal of Photogrammetry and Remote Sensing. 66 (2011) S2 S12.
[9] I. Puente, H. Gonzlez-Jorge, B. Riveiro, P. Arias, Accuracy verification of the Lynx Mobile
Mapper system, Optics & Laser Technology, 45 (2013) 578 586.
[10] H. Gonzlez-Jorge, I. Puente, B. Riveiro, J. Martnez-Snchez, and P. Arias, Automatic
segmentation of road overpasses and detection of mortar efflorescence using mobile LiDAR
49

Cuaderno Tecnolgico de la PTC

N 09 / 2014

data, Optics & Laser Technology, 54 (2013) 353 361.

[11] I. Puente, H. Gonzlez-Jorge, J. Martnez-Snchez, and P. Arias, Automatic detection of
road tunnel luminaires using a mobile LiDAR system, Measurement, 47 (2014) 569 575.
[12] M. Varela Gonzlez, H. Gonzlez-Jorge, B. Riveiro, P. Arias, Automatic filtering of vehicles
from mobile LiDAR datasets, Mesurement, 53 (2014) 215 223.

50

Road infrastructure design for optimizing sustainability

51

6. Acknowldedgements

Authors want to give thanks to the Xunta de Galicia (Grant No: IPP055 EXP44 and
EM2013/005; CN2012/269) and Spanish Government (Grant No: TIN2013-46801-C4-4-R;
ENE2013-48015-C3-1-R).

53

PLATAFORMA TECNOLGICA ESPAOLA DE LA CARRETERA (PTC)

Goya, 23 - 3, 28001 Madrid (Espaa)
Web: www.ptcarretera.es
E-mail:info@ptcarretera.es

En colaboracin con:

Con el apoyo de: