16 channels velodyne versus planar lidars based perception system for large scale 2d-slam

16 channels Velodyne versus planar LiDARs based perception system
for Large Scale 2D-SLAM
Nobili S.1 Dominguez S.2 Garcia G.3 Philippe M.4
Abstract— The ability of self-localization is a basic require-
ment for an autonomous vehicle, and a prior reconstruction
of the environment is usually needed. This paper analyses the
performances of two typical hardware architectures that we
evaluate in our 2D Simultaneous Localization and Mapping
(2D-SLAM) system for large scale scenarios. In particular, the
selected configurations are supposed to guarantee the possibility
of integrating at a later stage mobile objects tracking capabili-
ties without modifying the hardware architecture. The choice of
the perception system plays a vital role for building a reliable
and simple architecture for SLAM. Therefore we analyse two
common configurations: one based on three planar LiDARs Sick
LMS151 and the other based on a Velodyne 3D LiDAR VLP-
16. For each of the architectures we identify advantages and
drawbacks related to system installation, calibration complexity
and robustness, quantifying their respective accuracy for local-
ization purposes. The conclusions obtained tip the balance to
the side of using a Velodyne-like sensor facilitating the process
of hardware implementation, keeping a lower cost and without
compromising the accuracy of the localization. From the point
of view of perception, additional advantages arise from the
fact of having 3D information available on the system for other
purposes.
I. INTRODUCTION
An efficient and accurate solution to the Simultaneous
Localization and Mapping (SLAM) problem is the basic
building block for an autonomously navigating platform.
The perception system employed in the architecture plays
a fundamental role in determining the quality of the perfor-
mance. Indeed, the reliability of the map and consequently
the accuracy of the localization are highly dependent on the
measurements provided by the local sensors. Nowadays there
are several options when choosing the hardware architecture
that allows us to apply SLAM for positioning our vehicle.
The question arises when we have to choose which hardware
set-up is the most appropriate for our application given
some constraints about budget, ease of installation, preci-
sion, reliability against changing environment, versatility, etc.
Most commonly used architectures employ Light Detection
and Ranging (LiDAR) technology measuring at different
angles the distance to the surrounding environment. The
SLAM problem is well-known for its increasing complexity
in terms of accuracy, runtime and computational resources
1 2 4 Nobili S., Dominguez S. and Philippe M. are with IRCCyN, Institut
de Recherche en Communication et Cybernétique de Nantes, École Centrale
de Nantes, 1 rue de la Noë, 44321 Nantes, France
3 Garcia G. is with ECN École Centrale de Nantes, 1 rue de la Noë,
44321 Nantes, France
1 simona.nobili@eleves.ec-nantes.fr
2 salvador.dominguez@irccyn.ec-nantes.fr
2 Gaetan.Garcia@ec-nantes.fr
3 philippe.martinet@irccyn.ec-nantes.fr
required while covering long outdoor distances and map-
ping. We regard these aspects as particularly relevant for
precise localization of memory-restricted systems and we
propose a multi-map LiDAR-based 2D-SLAM solution. We
implemented an extended version of the GMapping Robot
Operating System (ROS) package. In particular, we adapted
it and integrated it into our architecture for being used in
Large Scale multi-map 2D SLAM. Our version allows to
start building a new map when required, saving previously
the map under construction. A map-manager is in charge of
deciding when a map must be stored and start building a new
one. Later, during the localization phase, the sub-maps are
loaded as they are required along the pre-recorded journey
and Monte Carlo localization techniques using a probabilistic
particle filter are applied to find the most likely position given
the map, laser scan and odometry measurements. Connection
points connect a sub-map with its neighbour and delimit
when a sub-map ends and a new one starts. The local
reference frame of a sub-map is normally positioned on a
connection point (See figure 1).
Fig. 1. We represent the vehicle’s path (from right to left) in a chain
of sub-maps. Each submap is connected to the previous and the next one
through connection points.
In this paper, we present experimental results obtained
in a urban context using two distinct laser-based hardware
architectures typically used in SLAM but the same software
for localization. In Table I, we summarize the main features
of both types of sensors. Our main contribution focuses on a
comparison between the performances of the two systems in
terms of map quality, localization accuracy and robustness
to temporarily static elements like parked vehicles.
The first experimental platform is an electric car Renault
Zoe ZE equipped with three LiDARs (SICK LMS 151)
placed at 50 cm from the ground level in a configuration

that guarantees a 360 Field of View (FoV) about the car
vertical axis. Specifically, two of the LiDARs are mounted
on the two front corners of the car in order to cover straight
and side views, whereas the third one covers the back side
view (See figure 2 and 4). An extrinsic calibration process is
performed to ensure that all the three LiDARs lie on a plane
as closed as possible to a common horizontal plane such that
the scans can be merged and given as input measurement to
the SLAM solver.
Sick LMS151 installed on
the front
Sick LMS151 on the rear
side
Fig. 2. The LiDARs Sick LMS151 are installed in strategical positions to
ensure a 360 FoV around the vehicle.
The second experimental platform is a vehicle Renault
Fluence equipped with a 16 channels Velodyne LiDAR
PUCK (VLP-16) placed some centimeters above the roof
surface and scanning 360 about the car’s vertical axis
(see Figure 3). This sensor is a 360 revolving-head 3D
LiDAR with 16 laser beams vertically separated along a
range of 30 with 2 of angular resolution. The ith laser,
after a full rotation, sweeps a cone in the 3D space. In this
context, we take advantage of the vertical FoV (from -15
to 15 with respect to the sensor reference frame) to infer
a 2D laser-scan information merging data belonging to a
vertical range between 1.8 and 2.8 meters from the ground.
In this way the scan measurements will not be influenced
by the most common moving elements, which in a urban
context are assumed to be cars, people or small objects.
This height also provides robustness to slopes and defects
of the road as the laser plane is less likely to intersect
the ground. Moreover, note that the decision of working
in two dimensions is justified by one main applicability
reason. From a practical point of view, 2D information are
sufficient for self-localization on a local flat map, as the car
moves locally in two dimensions, and are manageable in
the general case of restricted availability of computational
resources.
TABLE I
SENSORS’ FEATURES
- Field View Max Range Layers ⇡ Price
SICK LMS 151 270 50m 1 3000 $
Velodyne VLP-16 360 > 100m 16 8000 $
The remainder of this paper is organized as following. In
the next section, we present some of the most effective per-
ception systems currently employed for SLAM applications,
Fig. 3. The Velodyne VLP-16 is installed on the Renault Fluence ZE’s
roof surface.
along with some relevant state-of-the-art methods for large-
scale SLAM. In section III we explain how we obtain a 360
planar laser scan with a certain angular resolution where the
measurements are relative to the vehicle’s reference frame
in both cases. Additionally we explain how we generate the
ground truth, as well as, how we perform the comparison
between computed position and ground truth position. In
section IV we present the experiments performed on this
study and their purpose. In section V we present the results of
the experiments performed giving some partial conclusions.
And finally in section VI, we summarize the main points of
our results.
II. RELATED WORK
In the past, highly effective SLAM techniques have been
developed and state-of-the-art SLAM solvers are now able
to achieve good performances in terms of accuracy and real-
time processing (e.g. GMapping [1] and Hector SLAM [2]).
The first implementations of SLAM methods were based
on combined motion control and features observations with
an Extended Kalman Filter (EKF) [3], [4]. However, as
reported in [5], the solution to the EKF-based SLAM is
consistent in the linear-Gaussian case but diverges in general
situations. Subsequently, Rao-Blackwellized particle filters
have been introduced as effective SLAM problem solvers
under conditions of non-linearity. The approach proposed
in [1] and [6] uses a particle filter in which each particle
carries an individual map of the environment and treats
the consequent requirement of reducing the number of
particles. This algorithm is open source for the community
of researchers under the name of GMapping and is currently
employed for many SLAM-based applications. However,
the problem of computational complexity over large-scale
environments, of the order of tens of kilometers, has not
been directly addressed in this work.
Closely related to the solution we propose are hierarchical
SLAM methods to deal with large-scale applications. Atlas
[7] is a framework which builds a two-levels hierarchy
combining a local Kalman Filter with global graph
optimization. Similarly, Hierarchical SLAM [8] is a
mapping technique which uses statistically independent
local maps interconnected at an upper level through an

adjacency graph. Subsequent proposals employ independent
local maps with robot-centred representation [9], local metric
maps associated with topological places in a topological
map [10], submapping methods in which the current submap
summarizes all the information [11], local maps with feature
positions and final robot pose associated with a global map
without robot poses [12]. In [13] the authors present a
SLAM technique requiring a small memory footprint. This
feature makes this solution particularly suited for large-scale
problems. In the case of this paper, we present an adaptation
of the GMapping framework to deal with computational
complexity problems while covering long distances.
Of fundamental importance for a SLAM-based
architecture is the perception system. While quality
and accuracy of the sensors are basic requirements for
the reliability of the measurements, the costs in terms of
system installation effort and finances have to be taken
into account at conception time. During the DARPA Urban
Grand Challenge in 2007 [14], fully equipped autonomous
vehicles performed complex SLAM-based tasks to compete
in a 96km course in a urban area. Boss, by the Carnegie
Mellon University and General Motors Corporation team
[15], took advantage of three 3D LiDARs (IBEO Alasca
XT) mounted at about 50cm from the ground assuming
to cover a relatively flat road surface, processing each
layer independently and assuming infrequent changes in the
static environment. Junior, by the Stanford University team
[16], used a Velodyne HDL-64 mounted on the car’s roof
surface to perform 2D mapping and localization. Odin, by
the Virginia Tech team [17], used the coupled information
provided by three LiDARs (two IBEO Alasca XT and a
IBEO Alasca A0) mounted at about 50cm from the ground
level. In the last few years Google developed hardware and
software architectures for a self-driving car [18]. The heart
of the system is a Velodyne HDL-64 generating a detailed
3D map of the environment which requires high processing
capabilities.
In this context, the aim of our work is to provide a
comparison of the results obtained for the localization in
an unknown environment using two differently equipped
vehicles. In particular, we analyse the performances of our
solution to a large-scale SLAM problem in the case of
a (three) 2D LiDARs based versus a 16 planes Velodyne
LiDAR based perception architecture.
III. METHODOLOGY
The required inputs to the SLAM system are odometry
and planar laser scans, providing the information about the
car’s motion and the surrounding environment respectively.
Specifically, we generate 360 FoV scans with 0.5 of
angular resolution @ 10 Hz from each of the laser-based
perception architecture and odometry @ 100Hz from the data
collected by the OBD-II connectors and IMUs integrated
in the cars. For the sake of simplicity, we consider both
odometry and scans data with respect to the car reference
frame, placed at the center of the rear axis.
In order to provide an unbiased comparison, we run the
same SLAM system on both our vehicles. However, given
two distinct hardware architectures, the approaches adopted
to generate the scans and their quality vary. In the following,
we present the two different techniques developed to generate
laser scan information, first from the three Sick LiDARs and
then from the VLP-16.
A. Sick LiDARs
In the case of Zoe (Figure 4), we convert the individual
scan measurements into the 360 output scan by knowing the
pose of each of the sensors with respect to the car reference
frame. In particular, we determine the exact pose of each
sensor using a process of extrinsic calibration, fitting the
overlapping parts of the individual scans and then we convert
each of the measurements from the sensor’s local frame to
the car’s frame by simple reference frame transformation
(Equation 1).
Tscan = Tsensor ⇤ Tpoint (1)
where Tpoint is the transformation matrix of a point with
respect to the sensor’s frame, Tsensor is the fix transforma-
tion of the sensor with respect to the car’s reference frame
and Tscan is the transformation of the measured point with
respect to the car’s reference frame which is composed by a
translation vector {x, y, z} and a rotation matrix. From the
transformed position we can extract the angular position ↵
with the expression (2)
↵ = arctan
x
y
(2)
Depending on the scan’s angular resolution, for a given
↵, a unique corresponding index in the output scan vector is
given by (Equation 3).
i(↵) = round(
↵
↵
) (3)
where ↵ is the chosen angular resolution. In our case
↵ is 0.5 and ↵ 2 {0, 360 }.
On the other hand, we obtain the output range as
rangei =
p
x2 + y2 (4)
Notice that, for the same ↵ there can be more than one
measurement. In this case, we select the nearer point, i.e. the
one with smaller range.
B. 16 Planes Velodyne
In the case of Fluence (Figure 5), the VLP-16 is placed
above the roof surface. In particular, we compute the exact
position of the VLP-16 as 1.457 meters from the car’s rear
axis and 1.565 from the ground after a process of extrinsic
calibration. Similar to the three Sick LiDARs’ case, we con-
vert each of the measurements from the sensor’s local frame
to the car’s frame by simple reference frame transformation
(Equation 1). Notice that in this case the raw data provided
by the VLP-16 correspond to 3D measurements. Therefore,

Fig. 4. Sensor configuration for Zoe. In different colours, the coverage of
each planar LiDAR. In grey, the uncovered area. At the bottom, the profile
of the scans horizontal plane.
so as to obtain a planar scan information, we project the
points belonging to the vertical range {1.8, 2.8} meters from
the ground to a plane passing through the VLP-16 reference
frame center and perpendicular to its vertical axis. Once
this transformation is performed, the scan data type can be
identified by an index and a corresponding range value as
explained in the case of Zoe (Equation 3 and 4).
Fig. 5. Sensor configuration for Fluence. A 360 laser scan is obtained
from the 3D point-cloud generated with VLP-16. In the bottom, the profile
point-cloud range involved in the generation of the planar scan is shown.
C. Comparison with ground truth
The ground truth is generated using a Proflex 800 RTK-
GPS receiver that applies the differential corrections obtained
from a DGPS station located on the roof of IRCCyN build-
ing. The measurements provided by the RTK-GPS have an
error of less than 1 cm in position when in Fixed mode.
We estimate the orientation (heading), by computing the
direction of the movement.
We compare the position obtained by the SLAM system
at time t with the ground truth interpolated to that time.
The interpolation method used is through splines of position
coordinates with the time as independent variable.
IV. EXPERIMENTAL SET-UP
A couple of experiments have been performed to quantify
the localization accuracy and robustness of both the systems.
The precision of each of the results is evaluated comparing
the position of the car with a ground-truth generated using
the measurements provided by the high precision RTK-GPS
on the points where it is available. During both the experi-
ments, the vehicles travel in convoy, that is, one following the
other, in order to ensure the same environmental conditions.
A. First Experimental Set-up
For the first experiment, the vehicles travel for about 3
kilometers. The chosen trajectory covers the area around the
campus of École Centrale de Nantes, which is a standard
urban sector. Using the data recorded by both the vehicles,
we perform the task of map building. Later Monte Carlo
localization in these maps is performed. The aim of this
experiment is to compare the positioning accuracy of both
systems under the same conditions.
B. Second Experimental Set-up
A second experiment has been conducted in a highly
changing environment such as the parking area of the campus
of École Centrale de Nantes, where the position and num-
ber of parked vehicles constantly changes. We performed
recordings over two different days in order to ensure changes
in the temporarily-static elements (e.g. parked vehicles) met
along the path. In particular, the first day we recorded in
the early morning (when the parking was almost empty)
and the second day during the day-time (when the parking
was crowded). In this context, on the first day the vehicles
are asked to perform a task of localization and mapping.
Subsequently, on a second day, the vehicles have to localize
themselves along the same trajectory but with the maps
previously built. The aim of this experiment is to check the
robustness of each set-up against environmental changes.
V. RESULTS
A. First Experiment. Same trajectory, same environment
In figure 6 we show the accumulative and differential
histogram of the error resulting from localization under the
conditions detailed in Section IV-A, for both the cars.
As we can see the precision of Zoe’s localization
(equipped with planar LiDARs) is slightly better than Flu-
ence’s localization (equipped with VLP-16). For Zoe 95% of
the measurements have less than 0.6 meters of error while
for Fluence 95% have less than 0.7 meters. In this case,
Zoe performs better because, since the LiDARS detections
originally belong to a plane, the measurements are more
stable and less noisy. Instead, the fact that each VLP-16 scan

Fig. 6. Accumulative and differential histogram of positioning error
between both cars for the same trajectory and same environment. The
distance covered is about 3 Km. The coloured vertical line show the
maximum error obtained along the journey for the respective car.
results from the projection of a collection of points onto a
plane, causes the map not to be overall as well shaped as in
the case of the planar LiDARs.
B. Second Experiment. Same trajectory, highly changing
environment
Figure 7 shows part of the map built on the same area
by both systems. In the left image, the map built by the
VLP-16-based system doesn’t show short objects like the
cars parked, but the walls of the sourrounding buildings and
trees. In the right one, the map build by Sick LMS151-based
system show all the objects that can be seen at a height of
50 centimeter from the floor level. Just by looking at both
maps we can expect significant differences on the analysis
of the positioning accuracy.
Detail of a sub-map built
using the LiDARs VLP-16
Detail of a sub-map built
using the LiDARs Sick
LMS151
Fig. 7. Detail of the same working area seen by each LiDAR system.
Note that, while in the case of the Sick LMS151 the cars parked on the
sides of the road are visible while in the case of the VLP-16 only the static
environment like the walls of the buildings and the trees in the right side
are considered.
Figure 8 shows the accumulative and differential position-
ing error histograms for the case of the planar LiDAR-based
system. As expected the precision obtained for the second
day using the maps of the first day is worst than in the case of
the first day. However, an acceptable precision even under
highly changing conditions was not an obvious result. In
the first day 95% of the measurements lie under 0.4 meters
while in the second that percentage moves to 0.8 meters.
In this case we can assess that the common environmental
changes affect the precision but still more than 97% of the
measurements lie under 1 meter of error.
Fig. 8. Accumulative and differential histogram of positioning error for the
car equipped with three Sick LMS151 on the same trajectory but different
days where the environment has changed strongly. In coloured vertical lines
are marked the maximum error obtained for the respective day.
Figure 9 shows the accumulative and differential his-
tograms of the error in position between both days for the
case of the Velodyne-based system (Renault Fluence). As
expected, there is no significant difference on the precision
between localizing with the same maps in both days as, from
the point of view of the sensor, the environment has not
changed at the projected range of heights. This means that
this system is more robust against high environment changes
under heights below 1.8 meters than the case of the planar
LiDARs. Regarding the accuracy, in both days 95% of the
measurements have less than 0.6 meters of error in absolute
position.
Fig. 9. Accumulative and differential histogram of positioning error for
the car equipped with Velodyne VLP-16 on the same trajectory but different
days where the environment has changed strongly. In coloured vertical lines
are marked the maximum error obtained for the respective day.
VI. CONCLUSIONS
In this paper we perform a quantitative comparison, using
our 2D-SLAM localization system for large scale scenarios,

of two typical LiDAR-based hardware configurations: one
based on several LiDARs strategically installed around the
car at a low height, and the other one based on a single 3D-
LiDAR installed on the roof of the car. Both systems generate
a single 360 scan centered on the car’s reference frame
which is used for map-building and/or map-localization. The
qualitative results from studies conducted over our datasets
are summarized in Table II. In the same conditions and
with an static environment, the set-up using planar LiDARs
performs slightly better than with the 3D sensor as the
scans are less noisy on the range dimension. The positioning
error obtained is, in general, 10 cm. smaller in the first
case. The reason for being more noisy in the 3D sensor’s
case is because we are projecting onto the same plane a
range of heights (1 meter of range in our case), so that,
the probability of variability is higher than in the case of
one single height. Another conclusion is that the maps built
with VLP-16 data provide more robustness against common
medium term environment changes, like cars parked on the
sides of the road, describing mostly the static details, like
the walls of the buildings, trees or urban structures. This is a
desirable feature on map localization. Furthermore, there are
other advantages of using a single sensor regarding the sim-
plicity of installation, extrinsic calibration and general set-
up, speeding up considerably the set-up process. Regarding
the cost-effectiveness, the newly released sensor VLP-16 is
affordable for most of research projects in comparison with
using 3 outdoors planar LiDARs. Finally, in the case of the
3D-sensor the fact of having 3D information available on the
system opens a wide range of possibilities from the point of
view of the perception.
TABLE II
COMPARISON TABLE
- Planar LiDARs 3D LiDAR
Accuracy same environment " #
Robustness env. changes # "
Ease installation # "
Ease extrinsic calibration # "
Total cost " #
Provides 3D data No Yes
ACKNOWLEDGMENT
This work has been funded by the European Fund in
Regional Development FEDER-ROBOTEX trough the Pays
de la Loire regional development from june 2013 to october
2015. Thanks also to the ICARS team of IRCCyN.
REFERENCES
[1] G. Grisetti, C. Stachniss, and W. Burgard, “Improving grid-based
slam with rao-blackwellized particle filters by adaptive proposals and
selective resampling,” in Robotics and Automation, 2005. ICRA 2005.
Proceedings of the 2005 IEEE International Conference on, April
2005, pp. 2432–2437.
[2] S. Kohlbrecher, O. von Stryk, J. Meyer, and U. Klingauf, “A flexible
and scalable slam system with full 3d motion estimation,” in Safety,
Security, and Rescue Robotics (SSRR), 2011 IEEE International Sym-
posium on, Nov 2011, pp. 155–160.
[3] P. Moutarlier and R. Chatila, “An experimental system for incremental
environment modelling by an autonomous mobile robot,” in Exper-
imental Robotics I, The First International Symposium, Montréal,
Canada, June 19-21, 1989, ser. Lecture Notes in Control and Informa-
tion Sciences, V. Hayward and O. Khatib, Eds., vol. 139. Springer,
1989, pp. 327–346.
[4] R. Smith, M. Self, and P. Cheeseman, “A stochastic map
for uncertain spatial relationships,” in Proceedings of the 4th
International Symposium on Robotics Research. Cambridge,
MA, USA: MIT Press, 1988, pp. 467–474. [Online]. Available:
http://dl.acm.org/citation.cfm?id=57425.57472
[5] M. Dissanayake, P. Newman, S. Clark, H. Durrant-Whyte, and
M. Csorba, “A solution to the simultaneous localization and map build-
ing (slam) problem,” Robotics and Automation, IEEE Transactions on,
vol. 17, no. 3, pp. 229–241, Jun 2001.
[6] G. Grisetti, C. Stachniss, and W. Burgard, “Improved techniques for
grid mapping with rao-blackwellized particle filters,” Robotics, IEEE
Transactions on, vol. 23, no. 1, pp. 34–46, Feb 2007.
[7] M. Bosse, P. Newman, J. Leonard, M. Soika, W. Feiten, and S. Teller,
“An atlas framework for scalable mapping,” in Robotics and Automa-
tion, 2003. Proceedings. ICRA ’03. IEEE International Conference on,
vol. 2, Sept 2003, pp. 1899–1906 vol.2.
[8] C. Estrada, J. Neira, and J. Tardos, “Hierarchical slam: Real-time
accurate mapping of large environments,” Robotics, IEEE Transactions
on, vol. 21, no. 4, pp. 588–596, Aug 2005.
[9] J. Castellanos, R. Martinez-Cantin, J. Tards, and J. Neira,
“Robocentric map joining: Improving the consistency of ekf-slam,”
Robotics and Autonomous Systems, vol. 55, no. 1, pp. 21 – 29, 2007,
simultaneous Localisation and Map Building. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0921889006001448
[10] F. Ferreira, I. Amorim, R. Rocha, and J. Dias, “T-slam: Registering
topological and geometric maps for robot localization in large environ-
ments,” in Multisensor Fusion and Integration for Intelligent Systems,
2008. MFI 2008. IEEE International Conference on, Aug 2008, pp.
392–398.
[11] P. Pinies, L. M. Paz, and J. Tardos, “Ci-graph: An efficient approach
for large scale slam,” in Robotics and Automation, 2009. ICRA ’09.
IEEE International Conference on, May 2009, pp. 3913–3920.
[12] S. Huang, Z. Wang, G. Dissanayake, and U. Frese, “Iterated d-slam
map joining: evaluating its performance interms of consistency,
accuracy and efficiency,” Autonomous Robots, vol. 27, no. 4, pp.
409–429, 2009. [Online]. Available: http://dx.doi.org/10.1007/s10514-
009-9153-8
[13] B. Suger, G. Diego Tipaldi, L. Spinello, and W. Burgard, “An approach
to solving large-scale slam problems with a small memory footprint,”
in Robotics and Automation (ICRA), 2014 IEEE International Confer-
ence on, May 2014, pp. 3632–3637.
[14] (2007) Darpa urban challenge. [Online]. Available:
http://archive.darpa.mil/grandchallenge/index.html
[15] C. Urmson, J. Anhalt, H. Bae, J. A. D. Bagnell, C. R. Baker, R. E.
Bittner, T. Brown, M. N. Clark, M. Darms, D. Demitrish, J. M. Dolan,
D. Duggins, D. Ferguson , T. Galatali, C. M. Geyer, M. Gittleman,
S. Harbaugh, M. Hebert, T. Howard, S. Kolski, M. Likhachev , B. Litk-
ouhi, A. Kelly , M. McNaughton, N. Miller, J. Nickolaou, K. Peterson,
B. Pilnick, R. Rajkumar, P. Rybski, V. Sadekar, B. Salesky, Y.-W. Seo,
S. Singh, J. M. Snider, J. C. Struble, A. T. Stentz, M. Taylor , W. R. L.
Whittaker, Z. Wolkowicki, W. Zhang, and J. Ziglar, “Autonomous
driving in urban environments: Boss and the urban challenge,” Journal
of Field Robotics Special Issue on the 2007 DARPA Urban Challenge,
Part I, vol. 25, no. 8, pp. 425–466, June 2008.
[16] M. Montemerlo, J. Becker, S. Bhat, H. Dahlkamp, D. Dolgov, S. Et-
tinger, D. Haehnel, T. Hilden, G. Hoffmann, B. Huhnke, D. Johnston,
S. Klumpp, D. Langer, A. Levandowski, J. Levinson, J. Marcil,
D. Orenstein, J. Paefgen, I. Penny, A. Petrovskaya, M. Pflueger,
G. Stanek, D. Stavens, A. Vogt, and S. Thrun, “Junior: The stanford
entry in the urban challenge,” Journal of Field Robotics, 2008.
[17] A. Bacha, C. Bauman, R. Faruque, M. Fleming, C. Terwelp,
C. Reinholtz, D. Hong, A. Wicks, T. Alberi, D. Anderson,
S. Cacciola, P. Currier, A. Dalton, J. Farmer, J. Hurdus, S. Kimmel,
P. King, A. Taylor, D. V. Covern, and M. Webster, “Odin: Team
victortango’s entry in the darpa urban challenge,” Journal of Field
Robotics, vol. 25, no. 8, pp. 467–492, 2008. [Online]. Available:
http://dx.doi.org/10.1002/rob.20248
[18] (2011) Ieee spectrum: How google’s self-driving car works. [On-
line]. Available: http://spectrum.ieee.org/automaton/robotics/artificial-
intelligence/how-google-self-driving-car-works

16 channels velodyne versus planar lidars based perception system for large scale 2d-slam

More Related Content

What's hot (20)

Similar to 16 channels velodyne versus planar lidars based perception system for large scale 2d-slam (20)

More from Brett Johnson (20)

Recently uploaded (20)

16 channels velodyne versus planar lidars based perception system for large scale 2d-slam