Sagi Filin 
  
A LINEAR CONFLATION APPROACH FOR THE INTEGRATION 
OF PHOTOGRAMMETRIC INFORMATION AND GIS DATA 
Sagi FILIN, Yerahmiel DOYTSHER 
Department of Civil Engineering (Geodesy) 
Technion - Israel Institute of Technology, Haifa 32000, Israel 
filin.1@osu.edu, doytsher@geodesy.technion.ac.il 
Working Group IV/3 
KEY WORDS: GIS, Map Conflation, Map Revision 
ABSTRACT 
Knowledge provided by GIS data can alleviate many problems associated with object recognition from aerial imagery. 
However, as the scale of the GIS increases (for example a GIS database of topographic maps) positional disagreements 
between the data sets hamper the efficient utilization of this data. Presented here is a novel approach to the registration of 
GIS and photogrammetric data, based on local transformations according to linear features existing in both data sets. The 
results of applying this algorithm show great improvement in GIS accuracy and thus enable better utilization of common 
feature extraction techniques. The algorithm, being general, can be applied in its entirety or in part to different tasks 
concerned with integration and evaluation of several vector data sets. 
1 INTRODUCTION 
Autonomous extraction of geographic information from digital imagery is one of the greatest challenges of modern 
photogrammetry. However, this is a difficult task due to the complexity that characterizes natural scenes. The underlying 
problems that make it so complex relate to the amount of data, its variety both in terms of objects and object types, and the 
complexity of modeling and representing spatial relations between objects. Furthermore, current feature extraction 
algorithms are still far from being able to provide clean features for use by object recognition algorithms. Consequently, 
object recognition systems tend to focus on very limited tasks (road extraction, as a typical example, see Baumgartner et al. 
1997) in very limited environments (such as high altitude imagery or rural scenes) and even so have a very limited rate of 
success (Heipke et al. 1997). An alternative approach to overcoming these limitations utilizes existing GIS databases as part 
of its prior knowledge, thus reducing the uncertainty in the whole procedure. Integrating existing GIS data into the object 
recognition scheme is reasonable. Conceptually, it represents the existing prior knowledge about the data, thus reducing the 
uncertainty in annotating a given scene. Practically, it is a useless to detect road objects when the road's outline is already 
given in another database. Moreover, utilizing existing scene descriptions may help focus on more specific tasks such as 
map revision and improving the accuracy of the GIS, rather than concentrating on problems posed by the limitations of the 
feature extraction algorithms. 
An important issue that is usually overlooked in this regard is the inherent distortions between the GIS data and the imagery 
data. This phenomenon might not be as common in data arriving from large-scale mapping (cadastre for example) but it is 
widespread in medium to small scale mapping (topographic mapping for example), the data which is our concern. Here, 
cases of objects shifted tens to hundreds of meters away from their expected positions and distorted in respect to their 
original shape, are common, thus forming a large search space in the image domain for detection of objects. A common 
solution that is employed is to improve the registration between data sets. This is accomplished by either a global warping 
transformation or local, rubber-sheeting transformations, that achieve better accuracy. A rubber-sheeting transformation of 
the GIS features is usually based on using counterpart seed points between both data sets, mostly intersections of road 
networks, and performing local transformations according to the disagreement between them. Subsequently, after detecting 
counterpart seed points, they are triangulated to subdivide the plane into local regions, and then Piecewise Linear 
Homeomorphic (PLH) transformations are applied inside each triangle to transform the objects within. This widely used 
approach is very limited in truly coping with distortions between sets, since only the seed points are matched and the 
relative disagreements between the linear features shapes remain unresolved. Alternatively, we may utilize linear features as 
  
282 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.