by two
ct. The
uality of
'eometry
>xposure
nly the
ken into
obtained
with the
un,
and the
and the
of lines
baseline
and ij ;
and ij.
earch all
lence of
arget are
selected
t precise
nstrained
tified as
X, Yand
les. The
ted. The
ages and
the two
e image
ically by
e optimal
e further
1 in the
CTION
‘ormation
ignificant
zation of
plume of
1entioned
ge scale
>s, which
ocess of
ty of the
ment, an
proposed
or semi
automatic
e specific
objects. The user takes the responsibility of object recognition
and activates an object-directed algorithm for object
extraction. Guided by the user supplied information, the
scene- and image-domain knowledge can be incorporated into
an object model and, thus, a model driven automated
algorithm can be employed in the object extraction process.
The motivation of this strategy is to accommodate a friendly
environment of interaction between two important agents in
the system: human user and computer machine. A number of
object extraction algorithms have been developed. Aside from
the single-point based multi-image matching algorithm which
is described and used in the previous sections for object point
measurement, the following algorithms have been developed:
5.1 Road Centerline Reconstruction
Road centerline information marked by the lane markings is
very important to the generation of a road network
information system. It can be used to compute road inspection
parameters (longitudinal profile and surface deformation). A
global method to the automatic reconstruction of 3D road
centerlines from image sequences has been successfully
implemented to deal with the diversified appearances of lane
markings on roads (Tao et al., 1996). In this method, the
reconstruction is considered as a problem on "shape from
image sequences." The problem is to synthesize road
centerline information available from successive images into a
3D shape model. Firstly, a 3D physically-based shape model
of road centerlines is set up using the vehicle trajectory
determined by the combined GPS/INS navigation data. In
order to synthesize the constraint information coming from
object assumptions and image sequences into the model, the
model is defined as an active and deformable 3D curve (3D
snake). Cubic B-splines are employed to define this
deformable 3D curve model of the road. The physically-based
deformation mechanism has been incorporated into the model
such that the model can be progressively deformed under the
action of internal and external constraint forces. The extracted
3D points of road centerlines from image sequences can act as
external energy which forces the model to deform towards its
desired position. A novel feature extraction and matching
algorithm is developed to obtain these 3D points of road
centerlines from image sequences based on the exploited
stereo-motion constraints. Internal energy arises from
smoothness constraints representing the natural characteristic
of the shape of road centerlines. It maintains the a priori
assumptions about the shape of the model. Under a
combination of the actions of internal and external forces, the
model will be deformed incrementally towards the final state
in which forces from different sources are balanced. The
model resulting at the end of an input sequence represents a
3D road centerline shape. Various tests have demonstrated
that this method functions very reliably even in situation
where the road conditions are far from ideal (Tao, 1996).
5.2 Road Boundary Extraction
It is by no means easy to set up an unified model of road
boundaries dealing with a large variety of scenarios.
However, the features of road boundaries varies relatively
Smoothly. After comparisons of different algorithms, a
matching based boundary tracker is developed. If an starting
point along a road boundary is initiated by the user, the
235
tracker will follow this boundary automatically in the image
sequence. The least squares matching method is employed in
the following process. Considering the large distortion of road
boundaries in the image which seriously causes the matching-
based line following method to be corrupted, an object space
matching tracker is proposed. With the knowledge of the
known orientational parameters of images and the height of
camera station, the image window of road boundaries can be
rectified onto the ground plane. After the rectification, the
geometrical discrepancies along the same boundary are
reduced to a great degree and the object space based least
squares matching algorithm can be introduced for the
boundary following. In order to verify the tracking results
from single images, stereo matching method is also applied. In
this sense, when the matching tracker moves one step, the
least squares matching algorithm is applied along the
boundary within a single image and between two stereo
images. Finally, a Hough transfer algorithm is employed to
smooth the tracked points and eliminate blunders.
5.3 Vertical-linear Object Detection
Road corridor environments contain many vertical-linear
objects, such as traffic signs, power lines, telephone lines,
electronic poles, etc. The vertical line detection algorithm is
designed to assist the user to recognize and position such
objects. A further application of this kind algorithm is to
realize the automated recognition of road related structures
and infrastructure. The implementation of the vertical line
detection consists of four modules: (a) vertical edge filter; (b)
local linking of vertical edges; (c) Minimal Description Length
(MDL) edge segment representation; and (d) vertical segment
matching. Again, using the ground plane constraints, a
vertical low-pass filter is applied to detect the vertical edge
points whose angle is perpendicular to the normal of the
ground plane. A straight line controlled point linking method
is then used. Several rules about the knowledge of vertical-
linear objects, such as edge direction, contrast and shape, are
defined to restrict the linking process. To facilitate the line
based feature matching, the MDL method is applied to the
parametric representation of line segments. Consequently, the
line segment can be matched using the correspondence of the
feature set of line segments.
6. DATABASE GENERATION AND
DATA VISUALIZATION
Geo-referenced images yield 3D information by measurement
interaction. However, this system is not simply a survey data
processing platform that is limited to photogrammetric
functions. Instead, it is integrated with capabilities to allow the
construction of 3D objects. In recent research (Qian, 1996),
Objects are treated as semantically meaningful entities that are
more than lists of 3D coordinates, although at the low level
the 3D geometry of objects is constructed vertex by vertex.
The spatial relationships between the vertices are also
described. In the current implementation, a wire frame model
is used to describe 3D geometric structures. In a general 3D
information system, geometry is only one aspect of an object,
in addition to its rich descriptive information. Close-range
images give intuitive impressions of object types although
practically we need some additional information to give a
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996
