photogrammetric triangulation. Consequently, an integrated 
approach to multiple-image-based object measurement was 
developed (Tao et al., 1996 and Tao, 1997). 
b. Multiple-image-based semi-automatic object measurement 
The integrated GPS/INS navigation technology allows the 
collected image sequences to be georeferenced rigorously in a 
global coordinate system. Thus, the multinocular epipolar 
constraint is available from the image sequences with known 
ego-motion parameters. As shown in Figure 2, given a pair (i, j) 
of cameras and a physical point P, the conjugate epipolar lines 
L,j and Lji can be determined. A scene point P produces three 
pairs of homologous epipolar lines. When the image point (/,, I jt 
/*) forms a triplet of homologous image points, then I, is 
necessarily located at the intersection of the epipolar lines L tj 
and L ilc , respectively, defined by /, and I k . Therefore, the search 
for homologous image points between two images can now be 
reduced to a simple verification at a precise location in the third 
image. For instance, checking that {I I , I 2 ) form a pair of 
homologous image points consists of verifying the presence of 
1 3 at the intersection of L 31 and L 32 . The more the images used, 
the more the constraints that are available. 
Figure 2. Multinocular stereo geometry and multiple 
epipolar constraint 
The proposed multiple-image matching method consists of the 
following steps: • 
• Manually select an object point in the current left image 
using the mouse pointer. 
• Automatically snap a critical point close to the input point 
using a feature snapping technique. 
• Perform the initial image matching between the current left 
image and the right image along the corresponding 
epipolar line to find a set of match candidates. If no match 
candidate is found, use the forward left image or the 
backward left image in the image sequence forward-current 
or backward-current image matching. 
• Conduct consistency filtering to determine and validate the 
best match point from the match candidates using the 
multinocular epipolar constraint. 
• Calculate 3-D coordinates of the object point using 
multiple-baseline photogrammetric intersection. 
Several techniques such as, feature snapping, weighted cross 
correlation, correlation coefficient analysis, disparity range 
constraint, and two-way matching double-check, are also 
developed to further reduce the matching ambiguities. It has 
been tested that the matching successful rate is very high (96%) 
compared to other matching methods. Moreover, the finally 
obtained accuracy of 3-D object coordinates is greatly improved 
due to the use of multiple image geometry. 
4.1.2 Road Centerline Reconstruction 
a. Overview 
The management of vehicles and infrastructure requires a high- 
quality and up-to-date road-related spatial information system. 
Road centerline information has been extensively used to 
generate road network information systems. Such information 
can also be used to derive road inspection parameters such as 
the longitudinal profile and surface deformation, which are 
important indicators for road maintenance. The acquisition of 
up-to-date road centerline data using conventional field survey 
is fairly difficult due to the cost and logistical reasons. 
Since the image features of road centerlines are relatively 
distinctive, it is expected that automatic reconstruction of road 
centerlines from mobile mapping images can be realized. The 
first trial was performed by He and Novak (1992). They applied 
two methods to automate the reconstruction of road centerlines 
from a stereo image pair. The first one relies on the definition of 
the centerline by a certain pattern, e.g., by a bright line. Once 
the defined patterns are detected, they are approximated by an 
analytical function. The second one permits the user to define a 
starting point of a centerline on the screen. A line following 
algorithm is then employed to trace this centerline from the 
bottom to the top of the image until it disappears. Xin (1995) 
and Tao (1997) implemented similar methods in the VISAT 
system. Three line-following algorithms, namely direction- 
matrix-based line following, least squares line following, and 
polynomial edge fitting were tested. However, none of them 
performed superior than the others. It is observed that these 
methods can only detect and reconstruct the centerlines on the 
basis of a stereo image pair. The capability of continuous 
reconstruction of road centerlines from a complete image 
sequence is limited. 
Based on a solid theory, deformable models, an effective and 
very reliable approach was proposed and developed to address 
the issue of automatic and continuous reconstruction of road 
centerlines from long image sequences (Tao, 1996, and Tao et 
al., 1998). 
b. Automatic and continuous reconstruction of road centerlines 
from long image sequences 
In this approach, the reconstruction of road centerlines from 
image sequences is considered as a problem of ‘‘shape from 
image sequences.” The problem is to use road centerline 
information available from successive images in a sequence and 
road shape knowledge to update and refine a 3-D shape model 
of road centerlines (see Figure 3). Thus optimized road