2-5-8 
The two 3D point sets, derived from each epoch, are then 
analyzed for temporal correspondence. This analysis is still 
currently under investigation. A linear, three-dimensional, least 
squares similarity transformation is proposed, in which the 
second 3D point set is transformed onto the first. The computed 
transformation parameters then represent the relative motion 
between exposures. More detailed description can refer to 
Chaplin and Chapman (1998). Tests on simulated and real data 
have shown the viability of the technique provide that the 
solution is appropriately constrained. 
5. CONCLUSIONS 
The combination of computational vision and digital 
photogrammetry technologies is the principal methodology used 
throughout the thesis research. It is a key to the development 
and implementation of the three automated approaches. The 
thorough understanding of the imaging geometry and the 
extensive development of valid constraints constitute very 
important parts of our research work. A number of state-of-the- 
art computer vision methods, namely, model-driven feature 
extraction, multiple-image matching and shape from image 
sequences, are developed. 
Another methodology extensively applied is the human- 
computer collaborative semi-automatic scheme. An interesting 
aspect of the history of mankind is the relationship between 
human and computer. Although the great achievements have 
been made in recent 10 years, the realization of fully automatic 
feature extraction is still restricted especially in 
photogrammetry. Since we acknowledge that it is a long time 
before the automated feature extraction algorithm can be 
available as a production tool, and there is no much probability 
to separate the operator from the system, why should we not 
take the operator as a component of the system at the beginning 
of the system design? (Tao and Lin, 1994). A combination of 
the computational power of the computer and the interpretative 
ability and knowledge of the human brain has been 
implemented in the approaches of object measurement and 
extraction of vertical line features. It has been demonstrated that 
the use of this methodology makes the system work more 
reliably and efficiently. 
The following conclusions can be drawn based on the research 
results and real data tests: 
• The model-driven feature extraction methods have been 
proposed in the extraction of road centerline features and 
vertical line features. The geometric and photometric 
models of the specific features are established firstly. Then, 
the algorithms tailored to these models are designed. For 
example, in the extraction of centerline features, the 
algorithms of oriented edge detection, dual-edge detection 
and edge filtering by a new majority voting technique are 
developed. In the extraction of vertical line features, the 
algorithms of initial edge detection, edge thinning, 
vertically oriented edge detection, primary line grouping 
and secondary line grouping are implemented. • 
• The multinocular epipolar constraint has been successfully 
used to design an effective strategy of multiple-image 
matching. The matching ambiguities can be greatly 
reduced by using multiple-image matching associated with 
the weighted cross-correlation, the correlation coefficient 
analysis, the disparity range constraint, and the two-way 
matching double-check techniques. The method and its 
similar versions have been used in all these proposed 
approaches. It is seen that the method can be applied to not 
only point-feature correspondence, but also line-feature 
correspondence. More importantly, the accuracy and 
reliability of the 3-D reconstruction results can also be 
greatly improved using the multiple corresponding points 
or lines derived by the multiple-image matching method. 
• The concept and method of shape from image sequences 
are proposed and developed. This is, essentially, a closely 
coupled bottom-up and top-down reconstruction strategy. 
The whole reconstruction process is designed based on an 
interaction between a 3-D deformable model and updating 
information derived from the stereoscopic image 
sequences. This design leads to a robust result because the 
local and global shape of the model is controlled under the 
combinations of the actions of the internal energy (model) 
and the external energy (image). The synthesis of multiple 
constraints has been implemented by the least action 
principle, a physically-based mechanism for constraint 
combination. The advantages of the method are that 
constraints both from model assumptions and image 
sequences can be incorporated into a mathematical model, 
and the constraint combination can be solved by a 
Lagrangian motion equation. This approach represents a 
successful accommodation of multiple constraints from a 
mobile mapping system. 
• Image bridging is one of the new concepts in the 
development of mobile mapping systems. Current work 
involves the investigation of techniques for the automatic 
and robust establishment of correspondence between two 
3D point sets, and techniques for verifying the 
correspondence, for example, back-projection and 
matching in image space. Following this, sequential 
position and orientation updating procedures will be 
applied. 
6. THE FUTURE WORK 
• Image capture rate 
The stereo-motion cooperation based approach has 
demonstrated great advantages on the correspondence issue. If a 
high image capture rate is available, an improvement to a more 
robust matching can be achieved by using a fusion of feature 
correspondence and optical flow methods. 
• Image compression 
The fast image compression and communication techniques 
should be developed, due to the huge volume of image data 
generated. Data link from the mobile vehicle to an office GIS 
and data dissemination through internet/intranet should also be 
considered. 
• Multi-sensor information fusion 
It is believed that the automatic sensing ability will be greatly 
enhanced by using multi-sensor information, such as color, 
laser, and sonar. For example, if a laser ranging device is used 
and angled downward/forward the road, the elevation data of 
the road surface can be collected. This information can be 
further incorporated to improve the reliability of the automated 
methods. However, some issues have to be addressed, such as 
multiple information fusion, registration, management and 
computational complexity.