The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008
surface (PCRS) as the a priori knowledge. As shown in Figure 6,
the corrected projection is perpendicular to M\ and this
eliminates most of the restoration distortions in classical IPM.
5. LANE LINE SEGMENT EXTRACTION
At each survey point, the images from all the cameras are
captured at the same instant, based on distance or time, at a
series of points along the route using the VISAT Log data
acquisition module. The data acquisition module, VISAT Log,
can be configured on the fly to trigger the cameras for any
desired distance. In urban surveys, this distance is typically 3
to 5 meters intervals between image sets, which will ensure
complete panoramic coverage. On highways, 5 to 10 meters
intervals are typically used. At each survey point, the cameras
cover a wide angle of the surrounding area. The aim of LLS
extraction is to extract the lane lines within the covered area
from each image set captured at the same survey point. ARVEE
is adaptive to different camera configurations, which greatly
increase the flexibility. This convention is due to the effective
LLS extraction algorithm.
For each camera, the image is applied with a linear feature
extraction algorithm (Cheng Wang, 2002). All the lane line
associated linear features are filtered by the following
constrains:
(1) Shape constrain: the lane lines is supposed to be a solid or
dashed line with a certain width; (2) Colour constrain: the lane
lines usually are of white or yellow colour, and (3) Geometry
constrain: the road is a flat surface and the lane line is with a
small curvature.
The filtered out linear features from different cameras are added
into the lane-line-associated linear feature set (LALS). All the
elements in LALS are combined to establish the 3D model of
the LALS in the mapping frame. In this stage, more constrains
are introduced to determine is the location of the LLSs and their
central lines. These constrains are mainly from the
observational correspondence across cameras. The major
constrains include: (1) Space distribution correspondence: the
observations to the same LLS from different cameras should be
close to each other in PCRS. (2) Colour distribution
correspondence: the LLS should have similar colour
distribution in difference observations, and (3) Heading
direction agreement: all the LLSs should agree to the heading
direction of the road. At this stage, attributes that describe the
characteristics of each LLS are also extracted. These attributes
will later be used to classify the lane line type. For each LLS, as
shown in Figure 7, the local image is separated into three parts:
the lane line covered region C, the left neighbour region L and
right neighbour region R .
Motivated by common features used by human, ARVEE
utilizes the following features to describe a LLS: relative
position and orientation to the body frame centre; dashness (the
dash-shape of the LLS); colour distribution of region C, L, /?;
the relationship between the three colour distributions; and
texture features.
Figure 7: Local image segmentation for feature extraction
6. MULTI-HYPERTHESIS LINKING
Given the LLSs extracted from each survey point, the aim of
lane line linking is to join the LLSs through the whole survey
image sets to form a continuous 3D model of the lane lines.
Each lane line could include several to hundreds of LLSs.
Multi-hypothesis Analysis (MHA) has proved to be successful
in many applications, such as multi-target tracking (Gong,
2005). ARVEE utilizes a revised MHA algorithm to perform
the lane line linking. The MHA has three steps: hypothesis
generation; likelihood computation; and hypotheses
management.
All the LLSs are kept in a graph structure as shown in Figure 8.
Each node represents a LLS. For instance, node L tJ is the yth
LLSs in /th survey image set. The lane line linking develops the
edges connecting the nodes. At each survey point, the
hypotheses are the possible connecting configurations between
the ends of current maintained node to the nodes in the future
image sets.
Start
• • •
• False detection
Missed
End
• •
Figure 8: Graph structure for multi-hypotheses analysis
The hypothesise generation step first calculates the possibility
of the connections between the maintained graph nodes and the
nodes from the current image set. The maintained nodes include
the ending nodes of all the links in maintained hypotheses.
They are not necessarily from the previous image set since LLS
extraction may have misdetections. The connection probability
is computed as:
P co„ = W P P P + w dPd + W /Pf
where p p is the position closeness possibility between the new
node to the maintained note, p d is the direction similarity
possibility, pf is the feature similarity; and w p w d w f are the
weights of this probabilities. In hypothesises generation step, all
the possible connection configurations are added to hypothesise
list. Then, hypothesises with low connection possibility are
pruned out of the list.
Given the hypothesises (the connections configurations)
obtained from the previous processing, the likelihood
calculation step calculates the likelihood of each new
hypothesis. This step introduces the information of all the
maintained hypothesises, to bring in an overall view of the LLS
graph. The likelihood is calculated as:
likelihood
likelihood i =
+
Ê
W co„Pco„j + W ,rjP,rJj + WfeaPfeaj
n
Condition*
0 otherwise
where / is the current image set number, n represent the number
of objects in the current hypothesis. P conj is the connection
probability of the yth connection; P trjj is the smoothness
probability of the new connection to the former link; P feaj is the
feature similarity possibility. W con , W trj , and W feaj are the
weights to these possibilities. Condition* represents a set of
constrains that make a hypothesis possible. The constrains
include: no multiple connection; and no crossing connection.
These constrains express the nature of a possible lane line.
518
