Research Projects

In a CyberCity buildings are the most important part. In this case, they are fully phototextured and include parts such as roofs with chimneys and skylights, facades with balconies and bay- windows, etc. In the past roofs were modelled interactively on photogrammetric workstations or special programs were used with intervention of the human operator. The current research is oriented towards the automation of the data collection process so that large numbers of buildings of a city can efficiently be collected and regularly updated.

The roofs are being reconstructed from aerial images, the initial body of a building (building box) derives from known building footprints from a multi-purpose digital map, this is being extended in the vertical direction up to the roof. Roofs will be phototextured from aerial images, the building box (facades) will be phototextured after improvements with automated measurements of facades-detail from street level photography.

As an input dataset for the roof reconstruction we use aerial images with known orientation parameters, digital elevation models (DEM) and GIS-data (building footprints at the ground level and approximate elevation of the roofline from the multi-purpose digital map).

The creation of three-dimensional digital models of complex real-world objects is the task of this project. These models represent documents which have to be readable by a CAD-system for industrial use.
Our approach is based on high resolution digital images of the object from different view points to generate shape from stereo.

Algorithms and software from computer graphics for the photo-realistic rendering of complex scenes have reached a remarkable level of maturity over the last years. Little work has been done on evaluating the quality of such rendering algorithms on an objective quantitative or on a subjective qualitative basis. Such tests, however, would be very useful in order to rank rendering algorithms depending on the application domain and subsequently to automatically choose the algorithms most appropriate for the problem at hand.

Using an endoscope for the optical access to the combustion chamber of an engine makes it possible to study injection and combustion phenomena without compromising engine operation conditions.

Based on the endoscope technique the company AVL developed a new digital image capturing system, where a CCD camera instead of a high speed movie camera is used to observe combustion processes. This has great advantages in the area of image management and with regard to the possibility of online image processing. The disadvantage of using a CCD camera is that only one image per combustion cycle can be obtained. Stochastic cycle-to-cycle variations of the combustion process make it necessary to capture and analyze a statistically significant multitude of images of each crank angle.

With the confined number of images per crank angle a maximum of information of the combustion process should be gained with the help of image processing. One major goal of this project is to synthesize a movie sequence of the combustion process from images of different cycles. This image sequence should show a representative combustion process and the image to image transitions should be as smooth as possible.
For the analysis of images of a specific crank angle, methods are investigated to visualize cycle-to-cycle variations and to calculate or to select the most representative image.
From combustion images of Diesel engines the flame temperature can be determined by the two color method. For this purpose, the color CCD camera is calibrated using a tungsten striplamp to measure light density in two spectral ranges.

The purpose of this project is to successfully demonstrate the application of our general concept of information fusion in image understanding. The medical motivation lies in the extremely difficult diagnostic assessment of age-related macular degeneration (AMD).

Currently, six different Scanning Laser Ophthalmoscope (SLO) imaging procedures are used for a complete diagnostic data acquisition at the Vienna Eye Clinic.

Starting from this dataset, the correct diagnosis can be drawn by extraction and superposition of important retinal features. The features have to be extracted from several source images, but only information fusion of the different contents can yield the map of anatomical features and pathological changes.

Information fusion deals with the integration of information from several different sources. The necessity to fuse fuzzy information arises quite naturally in problems of image understanding because imprecision, uncertainty and ambiguity can be found at all levels, from the image itself to the results of high-level processing. In addition individual visual cues are often unreliable, even misleading. Thus integration of vision modules is necessary to obtain a reliable interpretation of complex images.

The Active Fusion approach followed in this research project has as its final goal the construction of a proto type system which is active in the sense that at each step of processing it can select and request the most promising next source of information and fuse the new results with the already gathered information. We expect Active Fusion not only to become a powerful new mechanism for the efficient control of an image understanding system but also to provide a way to deal with wrong, missing or ambiguous data. For example, the system is intended to be capable of actively choosing among different processing sequences and different views of the same scene to resolve ambiguities.

The research project Uncertain Information Fusion in Image Understanding pursues a thorough investigation of possibilities to incorporate the above ideas in an eigenspace based object recognition system.

In the steel-producing industry it is very important to control the quality of steels. At the moment the estimation is mainly done by human inspection of specially prepared steel surfaces with the help of light-microscopes. Each particular steel specimen is classified by assigning it to a reference chart, corresponding to the micro-structure of interest. For example, in the case of carbide distribution the reference chart consists of 28 classes arranged in 4 rows of 7 columns according to the 4 types and 7 degrees of the carbide distributions.

The purpose of this project is to automate the assessment of the micro-structure of steel specimens (e.g. the carbide distribution) by means of an active inspection system consisting of a computer-controlled light microscope and a workstation with classification and control software. The system yields a quantitative evaluation of the micro-structures, in contrast to the qualitative estimation of human inspection and consequently, more reproducible results. Moreover, an improved objective redefinition of the standard chart can be achieved.