18
The Figure 3 shows the same image as Figure1 as a product of a passive method. A known active method to
compensate the motion blur is the moving film of classical aerial cameras during the exposure time. The corresponding
passive method is the use of inverse filter algorithms. Another active method is the on-line control of CCD clock. At our
Institute the new aerial CCD- line scanner WAAC was developed. It has the capability of on-line control and we want
to present first ideas on its functioning.
3.1 On-line CCD clock control
Motion induced gaps in the raw image can not be corrected later and are the most important disturbance of raw
images. This disturbance can be quantified by the maximum distance between two following projected lines in the
rojection plane given in units of ideal ground pixels size in the plane. This gap can happen only, if the integration time
is less than the CCD cycle time. But this is the general case.
The changing v/H ratio and the fast changing attitude angles
destroy the equidistant sampling in the projection plane. The
ideal CCD clock AT,,, is that time, which is needed to move
over the distance B with the given velocity v. To avoid gaps in
the sampling we had two possibilities. The first one is
oversampling in any case by using a CCD clock time AT:
AT=K'AT,, AT, =(H-2,,) /v*f/dp k«1 (17)
The estimation of the needed k value has to be done before the
camera is switched on by evaluating the actual ratio of (H-z,_) to
the highest velocity in the projection plane. But the fixed CCD clock
and the varying horizontal velocities of each pixel will destroy the
inner geometry of the raw image in the same order as shown in
Figure 1. The only advantage is the avoiding of undersampling.
Figure 4 shows the maximum differences ‚in flight direction between
following lines in the projection plane in units of ideal pixels size (B)
(The used attitude data belong to the images in Fig.1. The
maximum difference occurs of course at the end of the projected
lines. The sufficient k value in this case was k < 0.5. The
consequence is a doubled raw data volume.
A better active strategy is to change k on-line. So we can avoid
even undersampling as oversampling in the projection plane and the
sampling is similar to the ideal lattice of lines and pixels in the
projection plane.
if FAT.) - fn > Ir, (+AT ) - 1 (1)
AT = [r,(+AT,.) - n],
else
AT = [ru trAT,,, = DIN ne (18)
Fig. 3 Geometrically corrected nadir image
r,,, is the component of r perpendicular to the last projected line in the projection plane and v, is the velocity component in the
same direction of one edge pixel. The on line control of AT is only possible, if we have predictions of r4, and va, .
This can be done using the methods of state estimation theory and Kalman filtering. We have to model the camera motion as
a stochastic process. The measurement noise and stochastic forces have to be introduced in the rigid body state equations It
is necessary to find stochastic forces, which have a normal distribution. A short analysis of angular accelerations shows, that
they could be modelled as normally distributed. Figure 5 shows for example the distribution shape of angular accelerations of
the roll axis. If the stochastic model coincides sufficient enough with the reality we get a prediction error which is normally
IAPRS, Vol.30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences", Zurich, March 22-24 1995
Fig.
10
Sanc
Part
Sche
the o
Schu
IAPF