LIDAR-BASED VEHICLE SEGMENTA TION
Á. Rakusz - T. Lovas - Á. Barsi
Department of Photogrammetry and Geoinformatics
Budapest University of Technology and Economics
H-1111 Budapest, Müegyetem rkp. 3.
Hungary
{rakusz.adam, lovas.tamas, barsi.arpad } @fmt.bme.hu
Commission II, WG 11/2
KEY WORDS: LIDAR processing, Object extraction, Algorithm comparison, Point cloud segmentation
ABSTRACT:
The paper focuses on a particular aspect of feature extraction from LiDAR data. To support transportation flow data estimation,
points reflected back from vehicles should be extracted from a LiDAR cloud. A simple thresholding can certainly provide a good
starting point to solve this task, but in order to achieve a robust solution there are several other tasks that should be addressed. First,
the road itself should be identified (actually continuously followed) to define the search window for the vehicles. Then, the surface
of the road must be modeled to obtain true elevation of the vehicle (which is measured in the normal direction of the surface). Once
the LiDAR points representing a vehicle have been obtained, at minimum the vehicle orientation should be determined such as travel
direction. This paper introduces a technique to accomplish the above mentioned tasks. The road is followed by the guidance of an
initial coarse centerline description. Then a preprocessing phase takes place, the point cloud is segmented to get the vehicle blobs.
The segmentation is based on standard image processing methods, such as histogram thresholding or edge detection techniques, both
methods are currently under consideration. In the next step, vehicle outlines are created using statistical parameters, such as standard
deviation of height values or height "texture" measures. The robustness of the process has been improved by using Delaunay-
triangulation to test slope measures. The newly developed method has been implemented in Matlab environment and provides
visualization tools for diagnostic purposes. The obtained results have proven that our algorithm performs well in effectively
extracting vehicles from LiDAR data that can contribute to the complex task of traffic flow information evaluation.
1. INTRODUCTION square grid, where the intensity comes from the height of the
point, certainly calibrated. In this paper we have used the
original sampled data and the resampled, interpolated form of it
(image). The visual control with images is also easier.
The methods were tested with two different datasets, with
diverse point density. One of them was acquired in July 2000
over the State Route US 35 (East of Dayton, OH), whilst the
other is about Toronto. Canada, in winter 2004. (See flight data
in Table 1) (Toth et al. 2003).
LiDAR stands for Light Detection And Ranging. Regarding the
data acquisition concept it is similar to radar, except it operates
with laser light. Flown with a helicopter or fixed wing aircraft,
eye-safe laser pulses are sent to the ground and their reflections
are recorded. Accurate distances are then calculated to the
points on the ground and therefore elevations can be determined
for not only the ground surface but the buildings, roads,
vehicles, vegetation and even something as thin as e.g. power
lines (Barsi et al. 2003), (Tovari 2002).
LiDAR technology provides a point cloud, in which all points Flight 1 Flight 2
have three coordinates and — in most cases — intensity values. Flying Height 470 m 660 m
The laser pulse reflects from the closest object, known as first (AGL):
pulse, (in this paper we have used only this). Average Ground 56.6 m/sec 58.5 m/sec
The greatest difference between LiDAR and other distance Speed:
measurement methods is the data structure. We have points Heading: 290 degrees (North- | 250 degrees
along a narrow strip, where we don't know exactly where the West) -
i Scan Frequency: 50 Hz 46 Hz
beam is reflected from, therefore, we cannot add any attribute
information to these points. In addition, when we use LiDAR
for transportation purposes, we have to adjust our calculation by
shortening against flight direction and elongating along it.
Therefore, our task is to develop methods for selecting.
separating and classifying points using different approaches.
One objective was to extract vehicle points for classification in
order to support transportation flow data estimation.
In LiDAR processing there are two completely different
possible approaches. If we don't want to lose information, we
have to use sampled 3D data points. But it is much easier if we
handle the data set as an image, after interpolating it to a regular
[56
Field of View (Half
Angle):
6 degrees
20 degrees
Laser Repetition
Rate:
10 kHz
70 kHz
Point density
1.5 points/m:
2.4 points/m*
Area
Route 35. Davton,
Ohio
Toronto, Canada
Table 1. Flight parameters
-
In the Ohio data set. 20-30 points are reflected from a passenger
car (and 40-60 from a truck) traveling along the flight direction.
Internat
In the Of
2004).
Figure !
view abo
EI
Fig
In this pa
segmenta
Image p
Surfacing 1
data, to a
Image
Labelin;
ncigb
Get LiD
Select ori
correspor
r
In Figur
(vegetatic
vehicles)
belonging
road, and
usable sw
If the cer
algorithm
located o
and the |
from the
lengths, r
the road
curve rad
should b
Combinir
the calcul
the match
the giver
caleulatio
should b
middle p
