2-5-4 
sensor mounted on a mobile mapping system could be treated as 
an active observer. Accordingly, we may state that mobile 
mapping systems provide an optimal experimental platform for 
research on active vision. On the other hand, the newly 
developed theories and methodologies of active vision offer us 
an invaluable tool to Tackle the challenge of automation in 
mobile mapping systems. 
3.5 Animate Vision 
It is argued that the active observer above defined is not truly 
active, but only a moving observer. Vision is not perception but 
a perception-action cycle. This is leading to another vision 
framework known as animate vision (Bajcsy, 1988; Ballard, 
1991). "We do not see, we look” represents the philosophy of 
this school. Under this framework, the process of vision is not 
considered alone, but as part of a global mechanism of an 
intelligent system, including cognition and motor processing. 
Current research work aims at developing active vision systems 
with great visual abilities, such as control of ocular parameters 
(e.g., aperture and focus), spatially-varying sensing, and gaze 
control (Abbott and Ahuja, 1990; and Burt, 1988). The control 
of the viewing parameters gives a stable and robust means for 
visual perception. The control of ocular parameters allows the 
system to maintain a suitable image quality against the 
degradations that often occur during the acquisition process. 
The control of gaze is commonly used in binocular camera 
heads (Ballard, 1991). This mechanism, called vergence, 
consists of bringing and maintaining the two camera axes at a 
specified spatial target position, fixation point. This permits the 
simplification of the correspondence problem. 
Animate vision further facilitates the computational process 
regarding 2-D correspondence and 3-D reconstruction to a large 
extent. We believe that animate vision theory and methodology 
will make a profound impact on the design and development of 
a new generation of mobile mapping systems, intelligent data 
acquisition and processing systems. 
4. AUTOMATED PROCESSING OF MOBILE MAPPING 
IMAGE SEQUENCES 
The combination of computational motion vision and digital 
photogrammetry technologies is the principal methodology used 
throughout the research. Great efforts have been placed on the 
development and employment of constraints from the mobile 
mapping system for design and implementation of the reliable 
information extraction methods. It is understood that the way to 
resolving an ill-posed vision problem is to exploit any possible 
sources of constraints. This is a key to the success of 
automation of image processing. The following constraints are 
derived and extensively applied to the methods proposed and 
developed: 
• Stereoscopic and sequential imaging geometry 
constraint, multinocular vision methods and stereo-motion 
image analysis techniques can be applied by using this 
constraint. • 
• Image geo-referencing constraint, rigorous epipolar line 
information and a direct image-to-scene transformation are 
available, since all the images have been geo-referenced in 
a global coordinate system. 
• Known vehicle ego-motion constraint: the viewer’s 
motion trajectory is determined by using GPS/INS 
navigation technologies. This information can be used to 
develop and optimize a road-network based information 
collection approach. 
Using the above constraints, methods for information extraction 
and image bridging from image sequences are developed. 
4.1 Information Extraction 
Since a huge volume of image data has been collected by 
mobile mapping systems, rapid and accurate extraction of 
features of interest from image data is highly desirable during 
'post-mission processing. The low efficiency of manual feature 
extraction is not compatible with rapid data acquisition by 
mobile mapping systems. It is also one of the major 
impediments in the development of on-line mapping or real 
time mapping systems. For these reasons, the emphasis of this 
research is placed on the development of methods for 
automatic, and accurate object measurement and feature 
extraction using mobile mapping image sequences. 
4.1.1 Object Measurement 
a. Overview 
The main task of mobile mapping systems is to map objects 
from images into a spatial coordinate system. The objects could 
be footprints of houses, street edges, centerlines, curbs, lane 
markers, manholes, culverts, fire hydrants, traffic signs, 
telephone booths, electric poles, etc. In order to calculate the 3- 
D coordinates of an object from images, at least two conjugate 
points in the images need to be determined. He and Novak 
(1993) applied an image matching technique to automate such 
an object measurement procedure. Since the orientation 
parameters of the cameras are known, the corresponding 
epipolar lines in a stereo image pair can be computed. Once a 
point in the left image is measured manually, the corresponding 
point in the right image can be determined by using the 
epipolar-line based image matching method. The area-based 
cross-correlation criterion was used in this method. 
In order to improve the reliability of image matching, edge 
features were used to constrain the matching results (Xin, 
1996). If a point measured in the left image is on an image edge, 
the corresponding point in the right image should be on an 
image edge too. However, this constraint is sensitive to the 
results of edge detection. This is a typical “chicken and egg” 
problem, is often encountered in computer vision research. 
Further improvement of the 3-D coordinate accuracy of object 
measurement was researched by Li et al. (1996). Their work 
focused on the use of multiple images to perform 
photogrammetric triangulation. Their results demonstrated that 
the final 3-D coordinate accuracy can be greatly improved if 
multiple corresponding points in image sequences can be used. 
It is required that multi-nocular point correspondence be 
established first. 
In fact, multiple images covering the same object are available 
in the VISAT mobile mapping system, that is, multiple 
corresponding points can be identified in the image sequences. 
The use of such redundant image information would be valuable 
not only for enhancing the reliability of image matching, but 
also for increasing the 3-D coordinate accuracy of