Road modelling using LiDAR-data

Abstract

The purpose of our master thesis is to locate and find a mathematical
model of the road surface from data obtained using LiDAR measure-
ments. A LiDAR is an optical instrument that generates a point cloud of
its surroundings by measuring distance and intensity. Some LiDARs also
generate additional metrics, but for our road model only the distance will
be used. However, we will discuss how the intensity can be used when
extracting road marking lines. The data used is from the KITTI Vision
Benchmark Suite [8] and was generated using a Velodyne HDL-64 LiDAR.
We have also simulated how our algorithm performs with lower resolution
LiDARs.
We will present two different methods and evaluate their performance
with regards to accuracy and speed. The methods will divide the data
generated from the LiDAR into triangular sections, and classify each sec-
tion as either road or not road. One of the methods will base its clas-
sification mainly on spatial information while the other will tone down
this aspect and instead filter results over time. To reduce complexity in
localization and computation, we have limited the filtering to a duration
of less than a second but the overall principles should work over longer
time periods. Due to a lack of pre-existing models of the road, we will
only be able to evaluate our methods visually. A goal has been that the
methods, with some alterations, should be able to run in real-time to as-
sist in autonomous driving. We find that both methods yield good models
of the road that reaches between 30 and 40 metres forward. The method
using time filtering shows the most promise, but the method with focus
on spatial dependency behaves better in some scenarios. Therefore, we
suggest combining the two methods in future work.

Popular Abstract

AUTONOMOUS DRIVING – – A WILD FANTASY OR A CLOSE REALITY?
Autonomous driving is one of today’s major
buzz words. Features like self-parking and lane
assistance that seemed like science-fiction only
10 years ago are today almost expected when
buying a new high-end vehicle. However, going
from driving assistance to full autonomy is still
a massive step. So, what needs to done?
The challenges are many but one area where
big strides need to be taken is perception. For
a human it’s easy to separate between a car
and a rock, or a patch of grass and a part of the
road. A computer has no such intuition.
Therefore, the vehicle is equipped with several
sensors, such as camera, radar, and LiDAR,
which provide necessary information required
for understanding its surrounding. The vehicle
needs a smart algorithm to make sense of the
data received from these sensors and that’s
where we come in. We have developed a
method that finds the road and creates a 3D-
model of the road surface using LiDAR-data.
This model can be used to know where to drive,
but knowing the shape of the road can also be
useful in other tasks. For instance, knowing the
distance to a car spotted by the camera, being
able to detect road marking lines, etc.
A LiDAR is an optical instrument that scans its
environment 10 times per second, each time
creating a 3D point cloud by sending out
infrared laser pulses and measuring the time it
takes for the pulse to return. A LiDAR will
therefore provide depth information,
something that a camera does not.
To be able to create a model of the road
surface, we apply a 2D-mesh consisting of
roughly 4000 triangles. We find which points
lies within each triangle and fit a plane to those
points. We compute the uncertainty for the
plane parameters and the inlier ratio, i.e. the
ratio of points lying within a given distance to
the plane. For the next scan we again calculate
the plane parameters, inlier ratio and their
uncertainties. The results from the different
scans are then combined, where planes with
high uncertainty will contribute less and vice
versa.
After five scans have been handled, each
triangle is classified as either road or not road
based on its compounded inlier ratio. The
planes classified as road are then connected at
the nodes of the triangles in order to get a
continuous model of the road.
Our method manages to model the road well
up to 30-40 metres forward, although false
positives sometimes occur. While the method
shows promise, there is still room for
improvement. For example when classifying
the triangles information from their
neighbours could be taken into account, this
way an isolated triangle would have stricter
requirements than one in the middle of the
road. However, we do believe that this method
is a good starting point for improving vehicle
perception, which in turn is an important step
towards full autonomy.

• author: Jungenfelt, Tove and Sevelin, Elias
• supervisor: Johan Lindström
• organization: Mathematical Statistics
• course: FMS820 20181
• year: 2018
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• publication/series: Master's Theses in Mathematical Sciences
• report number: LUTFMS-3351-2018
• ISSN: 1404-6342
• other publication id: 2018:E49
• language: English
• id: 8951555
• date added to LUP: 2018-06-20 09:20:32
• date last changed: 2024-09-24 14:53:44