AUTOMATIC ROAD VECTOR EXTRACTION FOR MOBILE MAPPING SYSTEMS
WANG Cheng 3 , T. Hassan b , N. El-Sheimy b , M. Lavigne b
‘'School of Electronic Science and Engineering, National University of Defence Technology, China - chwang_nudt@263.net
b Dept of Geomatics Eng., 2500 University Drive, N.W, The tfabbas@ucalgary.caof Calgary, Calgary, AB, Canada, T2N 1N4 -
tfabbas@ucalgary.ca, naser@geomatics.ucalgary.ca, mlavigne@amsvisat.com
Commission III: WG III/5
KEY WORDS: Mobile Mapping, Computer Vision, Road Geometry, Lane Line, Automatic Extraction, VISAT™
ABSTRACT: Land-based mobile mapping systems have yielded an enormous time saving in capturing road networks and their
surrounding. However, the manual extraction of the road information from the mobile mapping data is still a time-consuming task.
This paper presents ARVEE (Automated Road Geometry Vectors Extraction Engine), a robust automatic road geometry extraction
system developed by Absolute Mapping Solution Inc. (AMS). The extracted road information includes 3D continuous lane lines, road
edges as well as lane lines attributes. There are three innovations in this work. First, all the visible lane lines in the georeferenced
image sequences are extracted, instead of only extracting the central lane line or the nearby lane line pair. Second, lane line attributes
are recognized, so the output is a functional description of the road geometry. Third, the output is an absolute-georeferenced model of
lane lines in mapping coordinates, and is directly compatible to GIS databases. ARVEE includes four steps: First, extracting linear
features in each image. Second, extracting, filtering and grouping linear features into lane line segments (LLS) based on their
geometric and radiometric characteristics. Third, linking the LLSs into long lane lines 3D model using Multiple-Hypothesis Analysis
(MHA). Finally, classifying each lane line into a lane line type based on the synthetic analysis of the included LLSs’ features. The
system has been tested on large number of VISAT™ mobile mapping data. The experiments on massive real MMS data sets
demonstrate that ARVEE can deliver accurate and robust 3D continuous functional road geometry model. Full automatic processing
result from ARVEE can replace most of the human efforts in road geometry modelling
1. INTRODUCTION
Mobile Mapping Systems (MMS), provide an effective way to
collect georeferenced image sequences of the roads and their
surroundings. For instance, the VISAT™
(Video-Inertial-Satellite) (El-Sheimy, 1999) developed by
Absolute Mapping Solution Inc. (AMS).can be operated at road
speed of up to 100 km/hr and achieve abolute positioning
accuracy better than 0.3 m (RMS) for points within the field of
view of the images captured by the van. Mobile mapping has
yielded an enormous time saving in road network survey.
However, the manual extraction of the road information from
the mobile mapping data is still a time-consuming task.
Previous researches on lane line extraction mainly focus on the
traffic applications, such as traffic monitoring or autonomous
vehicle guidance (Ishikawa, 1988; Kenue, 1991; Jochem, 1993;
Chen, 1997; Beauvais, 2000; Paetzold, 2000; Yim, 2003; Li,
2004; McCall, 2004; Tai, 2004; Yue Wang, 2004; Hassouna,
2005; Jung, 2005; Lee, 2005; Choi, 2006), more details can be
found in (Kastrinaki, 2003). In summary, the constrains used in
lane line detection include: (a) the shape, the lane lines is
supposed to be a solid or dashed line with a certain width; (b)
the colour, the lane lines are usually white or yellow; and (c) the
geometry constrain, the road is flat and the lane lines are with
almost no horizontal curvature. Led by the application purpose,
and limited by the demand of real-time processing, these works
only concerned about lane lines that are close to the vehicle, and
all the results are described within local body frame coordinate,
or even simply within the image coordinate frame. In addition,
only vision sensors were exploited, and therefore, performances
are generally not satisfying at the situation of obscuration,
shadow or worn out painting. Few research works have focused
on lane line extraction in image sequences using,
georeferencing information from other sensing devices. For
autonomous vehicle guidance, Radar and camera fusion were
used to locate obstacle and lane line (Beauvais, 2000); location
sensing devices, such as GPS, were fused with vision in lane
lines following (Goldbeck, 2000; Jin Wang, 2005). In mobile
mapping, Tao (Tao, 2001) used georeferenced images form
mobile mapping image sequences to extract the 3D model of
central lane line. Roncella (Roncella, 2006) developed a
semi-automatic lane line extraction system and tested on
synthetic mobile mapping data.
Recently, we developed ARVEE (Automated Road Geometry
Vectors Extraction Engine) - a robust automatic road geometry
extraction system for the post processing of georeferenced
images captured by a land-based mobile mapping system. The
input of the system is the mobile mapping data, which includes:
georeferencing information, multi-camera panoramic images
sequence and sensor/system calibration parameters. The output
is the GIS-database-compatible road geometry information,
which contains 3D lane line model of all the lane lines visible
within cameras field of view together with line type/colour
attributes. The system works in a fully automatic mode, with no
operator supervision. The aim of the design is to introduce
computer vision techniques to do most of the road geometry
information extraction works in mobile mapping post
processing, and leave as less as possible work for manual
editing/correction.
There are three innovations presented in this work; first, all
the visible lane lines in the georeferenced image sequences are
extracted, instead of only extracting the central lane line or the
nearby lane line pair. The wider cover of each MMS survey
pass means less passes of the van to complete the whole survey.
This makes the MMS road survey more efficient. Second, the
