'et-
)3,
MATCHING IMAGES OF DIFFERENT GEOMETRIC SCALE
Franz Schneider and Michael Hahn
Institute for Photogrammetry, University of Stuttgart
Stuttgart, FRG
Commission III
0 ABSTRACT
In typical photogrammetric image processing applications scale differences between the images are usually small. There are
exceptions, however. In close range applications it can not be avoided that at least the image scales of some recorded objects differ
considerably from image to image. The same problem we meet in the stereo-image recordings of the MOMS-02 digital space-camera.
The pixel size related to the terrain surface is 4.4 m in the high resolution nadir channel and 13.2 m in the backward and forward
locking channels. For many digital photogrammetric tasks like point transfer, orientation, DTM reconstruction, etc., the problem of
image or feature matching has to be solved under such conditions.
The paper presents theoretical and experimental investigations into the matching problem with significantly different geometric image
scales. The special objective is to arrive at the highest possible precision by guiding the feature based matching by the high resolution
image. This implies multilevel and focusing techniques. The quality of the matching is assessed by the experimental investigations
with simulated and real images.
Keywords: Algorithm, Image Matching, Focusing, Resolution, Stereoscopic
1 INTRODUCTION ving vehicle, may evolve from a small blob to a detailed thistle
bush. This famous example is presented by Bobick and Bolles
The reason for this paper is a very practical one: Within the (1989) to explain, that recognition and tracking of objects needs
MOMS (modular electrooptical multispectral scanner) - project knowledge based systems and a suited representation space for
exists the problem of matching images which are significantly the description of an object at different scales. Sester (1990)
different in geometric scale. Images of the terrain surface are poses the question of treating the representation and reasoning
recorded by a three line scanner. Because of limitations of the problem in a multiresolution of a single image.
real-time recording capacity it has been decided to establish
different resolution channels to solve the stereo tasks of photo- The closeness in considering two images recorded with different
grammetry and mapping. Therefore in the panchromatric, high- geometric scale and two level of pyramidal representation of an
resolution stereomodule of STEREO-MOMS the pixel size in image are at hand. Assuming lowpass filtering of an image with
both, the backward and forward channel is 13.2 m, whilst in the a smoothing radius of o pixels and resampling a "new" image
downward channel the size of one pixel measures linearly 4.4 m is generated. The recorded and the derived "new" image has a
on ground (Ackermann et al., 1989). So naturally the questions scale ratio of o:1. The most simple procedure for the measure-
arises on how to exploit the more of information gained by the ment process would be to transform the higher resolved image
downward channel in combination with the lower resolved other to the scale of the lower one by lowpass filtering. Matching can
two channels. At first the scale difference of factor 3 between then performed with other images of similar scale. Of coarse by
the channels has consequences for all automatic measurement proceeding in this way the high quality information in the large
processes. This implies the selection of features in each channel scaled image is lost because of the smoothing operation.
as well as the establishment of feature correspondences. The
typical steps of photogrammetric data evaluation like point The procedure for intensity based matching of images presented
transfer, orientation and point determination (Ebner and Kornus, by Hahn (1990) allows to estimate geometry and radiometry
1991), DTM reconstruction (Hahn and Schneider, 1991), which transformation parameters, and furthermore the difference in the
are investigated in the MOMS-project, depend all directly or smoothness between the two images. The parameter used to
indirectly on the measurement process. Because the quality of measure smoothness differences is the Gaussian scale parameter
the measurement process propagates to all subsequent evaluation 0. This scale parameter of Gaussian smoothing is added to a set
processes it is of primer interest to investigate possibilities on of other parameters, i.e., it is also estimated by the procedure.
matching images of different scale. Similar problems we meet If the coarser resolved image is resampled to the sampling
in photogrammetric close range applications or in navigation density of the higher resolution image the effect of smoothing
with image sequences. The scale differences of objects imaged as mentioned above is still present. By application of the mat-
from different places may amount to 10046, in navigation di- ching procedure proposed by Hahn the smooth image and a
stinctly more in some tasks. In the latter case the appearance of stereo partner are matched. This implies, that the transformation
an object in an image sequence when approaching e.g. by a mo parameters between two images of different geometric scale are
295