ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision“‘, Graz, 2002
highest interest value. To ensure reliability, the maximum correla-
tion location may not be more than two pixels away from the ini-
tial feature on the search image. In addition, the correlation values
of the eight direct neighborhood pixels around the maximum
should also be locally large indicating the center pixel is on the hill
of a correlation surface. The 3 x 3 local correlation hill is repre-
sented by a quadratic two-dimensional polynomial with the maxi-
mum of the polynomial as the final solution of the correlation. The
O for correlation is better than half a pixel. To further improve,
least-square matcher (LSM) [Gruen and Baltsavias, 1987] is fol-
lowed. The same template camera is used. LSM sequentially
matches the template with every conjugate image patch instead of
simultaneously applied to multiple image. This way, the matcher is
separated from the engine that controls the matching legislation.
The correction of LSM is only allowed within one pixel for high
reliability. The ¢ for LSM is less than two tenth of a pixel.
Though this combination of correlation and LSM algorithms cre-
ates an expensive area-based matcher, it still has a high perfor-
mance due to the quality of the input TPs and the small search
windows.
4. EXPERIMENTS
4.1. Pre-launch Tests
In prelaunch development, an orbit of simulated MISR imagery
that runs through the north America continent was created, based
on similar bands of Landsat Thematic Mapper (TM) imagery
along with a registered DEM. The simulation process is described
in [Lewicki et. al., 1994]. The TP detection algorithm was imple-
mented and tested with the simulated image. The results were
applied to SBA for evaluations against various orbit error models.
Due to the dynamic errors in the simulated orbit, SBA requires 20
well distributed TPs per 512 image lines. SBA also shows cluster
of TPs at both sides of the swath are not necessary. TP detection
was configured accordingly. Figure 2 illustrates the number of
detected TPs per SOM block along the swath. A MISR ground
path on SOM projection is reported on 180 blocks from north to
south. Each SOM block is 140.8km long in the along-track direc-
tion and contains 512 lines of projected image. The land coverage
of the simulated MISR imagery is from block 51 to 77. Note the
TPs in the start and end blocks are less since they are partial
blocks. The overall TPs could be increased if configured to. On
average, there is over one TP detected per TPC per MISR camera
regardless surface type, with performance slightly better on hilly
terrain. The detection is slightly weaken on the most oblique D
cameras, covering only 8096 of the grids set by initial match.
Generic TP detector was intensively tested with the simulated
MISR imagery that cover different surface types and are resam-
pled from LandSat imagery of different times. It worked effec-
tively and consistently if initial match would supply local image
patches with over 30 pixels overlap in both dimensions. Though
most parameters are configurable, the system performed well with
the defaults in all tests. On average, there are 15-20 interest points
detected on a local image patch, and among them 10-15 interest
points are matched in feature base consistently (MISR D cameras
are likely on the lower end). The experiments also show that the
detection of invariant interest points over multi-image are critical
and the relational-based matcher is very reliable.
The accuracy of the matching was assessed in three ways. First,
TP symbols were superimposed on imagery and visually examined
by zooming into the pixel level for measurements. Second, “truth”
TPs were simulated for all detected TPs, by projecting from one
precisely matched camera down to the surface DEM, then back-
ward up to the rest of cameras. The projection used the simulated
navigation data with no added errors, in contrast with the “mea-
sured” navigation data used in TP detection and SBA. The mean
differences between the detected and simulated TPs are within 0.2
pixels for LSM points, 0.4 pixels for correlation points, and about
2 pixels for feature TPs, all with a small sigma. Finally, SBA
would evaluate the residuals of the TPs after trigulation and com-
pare with the errors added in the orbit model. The results verified
the uncertainties of various types of matching and the sufficency of
TP detection with respect to the need of triangulation.
60
T T T
"p27.orbitt measured.51.77.3.test.block tp" ©
40r ©
Figure 2. Number of detected TPs per SOM block.
4.2. Post-launch Operation
TP detection was tested with MISR data after Terra launched in
late 1999. With little change, it was set into production at an aver-
age rate of processing one day-side orbit per 45 minutes and run
automatically upon data coming. Figure 3 is an example of TP
detection on MISR orbit 2578 over south Greece for camera Df
and Aa, where the red symbols are TPs matched with LSM, the
yellow ones are matched with correlation, and the green ones are
feature matched. Note that image Df has less match as the camera
looks through a long path in the atmosphere and image becomes
fuzzy. Quality assessment statistics, such as the plot shown in Fig-
ure 2 as well as other SBA feed backs, are created automatically
for each orbit production. The overall production ended within a
few months and successfully covered all Terra paths. Though there
are a few paths with very few TP detections, the investigation
showed they all due to the imagery is too cloudy according to the
MISR cloud detection standard.
4.3. Adoption to AirMISR
The pushbroom AirMISR flies on 20km above the surface. Com-
paring with MISR, both its position and orientation are unstable,
causing large image distortions. Generic TP detector is applied to
imagery resampled on the ground using initial navigation. When
initial navigation provides near 30 pixels overlaps, generic TP
detector works fine as to MISR, regardless the much worst image
distortions. When image pointing is worse, initial match would
A - 428
