International Archives of the Photo;
erammetry, Remote Sensing and Spatial Information Sciences
. Vol XXXV, Part B3. Istanbul 2004
C)
Example frames from thermal image videos with
homography estimation overlaid: a) Video 1, oblique
with very small field of view; b) Video II, oblique
with still small field of view; c) Video Ill, forward
looking, low flying and strong rotations.
For Video I there was a ground-truth file containing pose
estimates obtained by a priori GIS/INS recordings and posterior
back-section with geo-referenced building models. This enabled
the estimation of the projective distortion matrix and systematic
offset like outlined in Sect. 2.5. The same distortion parameters
were used for Video II. Video Ill was regarded as distortion-
free (but with camera rotations). The outlier-rate and the
standard deviations for inliers where estimated. Deviations are
given in relation to the length of the measured vectors. Outliers
are usually defined as having more than 100% deviation (except
Video II with rotation, were 500% were used for f).
Video I Video II Video III
Outlier-rate
Rotation-included 100% 71% 32%
Deviationt . | 385% 43%
Deviationn — | ——— 64% 41%
Deviation R= | -——— 5% 10%
Outlier-rate
Rotation-free 58% 16% 31%
Deviation t 78% 38% 56%
Deviation n 69% 39% 48% r
Table 2. Deviations and outlier-rates
4.2 Conclusion
Particularly, if long focal lengths are used the estimation
accuracy for the camera pose that can be achieved through full
homography decomposition may be rather poor. With Video I it
turned out a complete random number generator. Often INS-
sensors are mounted on the same platform anyway that will
give high precision. In these cases a rotation-free decomposition
is favoured with the homography being a homology or an
elation. The clation case is not exception because aircrafts are
often operated at level that means with the epipole on the
horizon. Therefore the eigen-space decomposition is not
recommended. Instead the rotation free case allows estimating
the epipole directly from the best sub-set of correspondences.
Subsequently the plane parameters can be inferred linearly and
unambiguously from the decomposition. Pose estimation by
decomposing the homography of the image flow measures
velocities over ground in relation to flight altitude and the
absolute nadir direction (at least if the scene plane is levelled).
It may complete other sensors like INS giving accelerations and
rotations, GPS giving absolute locations, speed sensors giving
speed in air, altimeters giving absolute height and magnetic
compasses giving geographic direction. Special care has to be
taken for non-projective distortions of the camera and lens,
particularly with thermal cameras of the non-focal-plane-array
type. Such distortions may well be misunderstood by the
decomposition as rotations. It is desirable to calibrate a
systematic offset from comparing the system to ground-truth.
References
Ben-Ezra, M., Peleg, S., Werman, M., 1999. Real Time Motion
Analvsis with Linear Programming. CVIU, Vol.78, pp. 32-52.
Beutelsbacher, A., Rosenbaum, U., 1998. Projective Geometry.
Cambridge Univ. Press, Cambridge.
Faugeras, O., 1993. Three-Dimensional Computer Vision. MIT
Press, Cambridge, Mass,.
Fischler, M. A. Bolles, R. C. 1981. Random sample
consensus: A paradigm for model fitting with applications to
image analysis and automated cartography. Comm. Assoc.
Comp. Mach., Vol. 24, pp. 381-395.
Foerstner, W., 1994. A Framework for Low Level Feature
Extraction. In: Eklundh J.-O. (ed). Computer Vision - ECCV
94. Vol. II, B1, pp. 383-394.
Hartley, R., Zisserman A., 2000. Multiple View Geometry in
Computer Vision. Proc. Cambridge University Press,
Cambridge.
Holland, P. W., Welsch, R. E., 1977. Robust regression using
iteratively reweighted least-squares. Comm. Statist. Theor.
' Meth., Vol. 6, pp. 813-827.
Jurie, F., Dhome, M., 2002. Real Time Robust Template
Matching. BMVC-2002, pp. 123-132.
Michaclsen E., Stilla U., 2003. Estimation of vehicle movement
in urban areas from thermal video sequences. 2nd
GRSS/ISPRS Joint Workshop on Remote Sensing and data
fusion on urban areas, URBAN 2003, IEEE, pp 105-109.
Michaelsen E., Stilla U., 2003. Good sample consensus
estimation of 2d-homographies for vehicle movment detection
from thermal videos. In: Ebner H, Heipke C, Mayer H, Pakzad
K (eds) Photogrammetric Image Analysis PIA'03. [International
Archives of Photogrammetry and Remote Sensing. Vol. 34, Part
3/W8, pp 125-130.
Tor P HS, Davidson C., 2003, IIMSAC: Synthesis of
importance sampling and random sample consensus. 1EEE —
PAMI, Vol. 25(3), pp. 354-364.
Sawhney, H. S., 1994. Simplifving motion and structure
analysis using planar parallax and image warping. ICPR 94,
IEEE-Press, Los Alamitos, Ca., Vol. I, pp. 403-407.
Stilla, U., 1995. Map-aided structural analysis of aerial images.
ISPRS Journal of photogrammetry and remote sensing, Vol.
50(4), pp. 3-10
KEY WORDS
matrix
ABSTRACT:
For the past fifteen
vision communities.
fundamental tasks. \
in its various orient:
features often corre:
correspondence bet
arbitrary geometry)
resection problem, i
so far. Two addition
most basic photogr:
literature. This pape
features. We restrict
ICP algorithm and p
L
Line Photogrammet
of research for aln
researchers have a
features instead of
Some of their centra
I. In many typical
more reliably than [
urban and man ma
(Habib, 2001) 3.
industrial metrology
point features, thus :
automating the inv
Matching linear fe
matching point featu
This paper presents
determining the rela
proposed here are b
instead of conjugate
In the resent years \
more digital phc
automatic processe
attracted a large bo
computer vision €
automatic aerial tri:
entities for determi
solutions for relativ
Hough search techn
2001). Solutions fo
linear features, nam
been introduced by
