3d point generation which show a sufficient number of
homologue points and a big stereo base within the sequence. As
only small parts of a building facade are visible within one
image, a direct matching of edges extracted from the images and
a given building model cannot be performed directly. An
example of a recorded sequence is shown in figure 1. The
position of the camera was recorded with GPS and, for quality
measurements from tachymeter measurements from ground
control points.
Figure 1. Example images from one sequence along a building.
3.1 Feature Tracking
In a first observation, the number of features per image and the
number of images in a sequence a feature can be tracked is
investigated. Figure 2 shows an image of a sequence with SIFT
features detected in two images and the movement of the
features from the first image to the second.
Figure 2. IR image with selected SIFT features, that have
correspondences in the following image. Arrows show the
moving direction of the points and numer of pixels they move.
For small distances between the images, Foestner points and
SIFT features show almost the same number of homologous
points in two images. With a bigger distance of the images, the
number of homologous points from Foerstner points decreases
faster than the number of homologous points from SIFT
features. For sequences of several images, this decrease is much
smaller but also shows a better performance for SIFT features.
Table 1 shows the decrease of the mean number of homologous
points with the distance of images in the sequences. For
comparison reasons the number of features for the first image
was set to 100 for both feature detectors. For selected
sequences, the features have been tracked manually to see the
decreasing number of features due to features running out of the
image.
Distance in frames / 1/ 10/ 50 / 200 /
seconds 0.02 0.2 1.0 4.0
Manually tracked 100 916 1832. 1517
Foerstner point single pair | 99.5 89.7 | 47.5 (243
SIFT feature single pair 99.4 89.8 | 614 | 385
Foerstner points with 10 | 99.5 89:7 155.8 1 343
frames distance step
SIFT features with 10 | 99.4 89.8 | 645 | 42.4
frames distance step
Table 1. Decrease of the mean number of homologous points
with the distance of the images for Foerstner points and SIFT
features in the sequences.
3.2 Orientation of Image Sequences
The homologous points are used for the calucaltion of trifocal
tensors as introduced by Mayer (2007). The bundle adjustment
is additionally given the obervations of the position of the
camera of every image. The resulting oriented image sequence
is used to derive 3d coordinates of the homologous points. The
generated 3d point cloud of the SIFT features and the estimated
camera position for every image of the sequence can be seen in
figure 3. The structure of the facades is already visible. Most of
the points are located in the edges of the windows and grouped
in lines. The variance analysis of the bundle adjustment of the
relative orientation of the image sequences shows smaller errors
for SIFT features compared to Foerstner points due to the weak
edges in IR images and mismatches in window regions for
Foerstner points.
Figure 3. Point cloud of one image sequence along a group of
facades. The squares are representing the estimated camera
positions.
3.3 Matching with the building model
The 3d point cloud generated in the image sequence orientation
step is now matched with the given building model (Fig. 4). A
grouping of the points is done before the matching to remove
non façade and wrong points ie. of trees. The local
neighborhood of every point is analyzed to derive an estimated
plane the point is on and its normal. Every point is assigned to
the surface with the smallest distance and similar normal
direction. Points with a distance or normal direction that differs
beyond a threshold for all facades are rejected. The remaining
points are now used for the least squares matching with the
facades of the building model.
