Keywords: knowledge based system, videometric system, interest operator.

In science and industry (for example in architecture, medicine, or construction), highly accurate 3D representations of objects are required. A great variety of optical 3D measurement techniques like laser scanners, photogrammetric systems, or image based measurement system is available to achieve this need. Compared to laser scanners, imagebased systems measure objects with higher accuracy. Compared with photogrammetric systems, they can be easier to use for online measurement processes. This will especially be the case, if the measurements can be performed with a high degree of automation. Image based measurement systems perform their measurements with or without targeting. Some applications like, e.g., the monitoring of buildings definitely require measurement systems without artificial targets because they would highly disturb the architectural impression. A method to replace these targets is to use the texture on the object surface to find "interest points" by interest operators. The disadvantage of such systems is the need for a well trained "measurement expert" having special skills and experience to properly operate the complex system. In a complex measurement system with many algorithms for image processing and point detection, the selection of suitable algorithms, their order of application, and the choice of input parameters is a non-trivial task. To provide automated support, a knowledge based approach has been chosen for representing the knowledge necessary for this decision making, allowing for a declarative and modular representation of a decision policy together with easy extendibility. Extensive experimentation showed that the algorithms can be chosen and combined on the basis of parameters extracted from the image (low-level feature extraction). This is done by a special image analysis process [2].

A necessary precondition for the successful application of algorithms for finding interesting points is the "quality" of the image. It is often required to improve the visual appearance of an image. This can be done by image preprocessing and enhancement processes. Furthermore, flexible image processing makes the measurement system more independent of variable illumination during image capturing. The aim of the developed knowledge based image processing system is to select, on the basis of extracted image features, a single algorithm or a combination of algorithms for image processing in order to improve the image for the subsequent application of interest operators. At critical processing steps (e.g., edge detection, median filtering), the user can overrule the system decision.

The second component of our system is the knowledge based point detection by means of interest operators (IOPs). There exist many IOPs; however, no IOP is suitable to find all desired points. For this reason, we have implemented in our system different IOP algorithms (Förstner operator [1], Harris operator [3] and the Hierarchical Feature Vector Matching operator [4]). The choice of one or more suitable algorithm(s), their combination and parameter(s) is made by the KBS. The knowledge to be included in this part of the knowledge base was obtained by theoretical considerations and extensive experiments. We have used for the evaluation of interest operators several methods: visual inspection, ground-truth verification on the basis of good and bad areas, and a new evaluation method by means of distances between sets of interest points. Like the image processing algorithms described above, a combination of suitable IOPs is selected together with its parameters. Operator parameter values are specified on the basis of image features and features collected by user-queries.

In spite of choosing suitable image preprocessing algorithms and suitable interest operators, the number of detected points is often too high. Therefore, the next step has to be the reduction of the interest points by a special point filter. The filtering process is done by means of two methods: point filtering on basis of defined rules (knowledge based) and interactive point filtering (user based).

It is notable that we have conducted extensive experiments with the KBS on about 120 pictures showing different kinds of buildings. The system yields good results and shows a reasonable performance. To the best of our knowledge, the described system is the first "intelligent system" for such a complex measurement application. We report on the architecture and functionality of the respective knowledge based system, its development stage and the promising results obtained in experimentation.

1

Förstner, W. and Gülch, E., A Fast Operator for Detection and Precise Location of Distict Point, Corners and Centres of Circular Features. In: ISPRS Conference on Fast Processing of Photogrammetric Data, Interlaken, 1987.

2

Haralick, R.M. and Shapiro, L.G., Computer and Robot Vision, 1st Edition, Addison-Wesley Verlag, New York, 1993.

3

Harris, C. and Stephens, M., A combined corner and edge detector. In: 4th ALVEY vision conference, Matthews (ed.), University of Manchester, 1988.

4

Paar, G., Rottensteiner, F. and Pötzleitner, W., Image Matching Strategies. In: Kropatsch and Bischof (eds.), Digital Image Analysis, 1st Edition, Springer Verlag, Berlin/Heidelberg/New York, 2001.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £80 +P&P)