ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002 
  
  
parameters for decentering distortion and the affine 
transformation for scale differences between axes turned out 
to be insignificant. The radial symmetric component of the 
lens distortion is shown in Fig. 9. Clearly, this camera 
exhibits quite significant distortions towards the edges of the 
image sensor (on the order of 250 microns or 30 pixels). 
However, the extent of the radial symmetric distortion over 
the typical image measurement area is only about a few 
pixels or about one cm at ground scale. In the next step, 
images were acquired in a cloverleaf pattern over the ground 
target area by the camera installed in the mapping vehicle. 
Results from the triangulation with self-calibration using the 
AeroSys program have delivered virtually identical camera 
calibration parameters. 
Radial symmetric distortion 
EN 
  
M 
[mm] 21 28 {mm} 
Figure 9. Radial symmetric distortion. 
5.2 Boresight calibration 
Boresight calibration is referred to as the estimation of the 
linear and angular components of the transformation between 
the camera and the INS body frames. It is usually 
accomplished through a mathematical comparison of the 
aerotriangulation (AT) solution and an independent GPS/INS 
solution for the exterior orientation parameters. The boresight 
calibration of the system presented here has been performed 
at the OSU West Campus target area after the hardware was 
installed in the mapping van and thereafter for stability 
check. The sensor configuration is shown in Figure 10; the 
GPS antenna is approximately above the IMU center. 
  
Figure 10. Sensor geometry. 
A - 366 
In each boresight calibration session, a total of three-four 
independent EO solutions were determined, based on several 
independently collected sets of three images each. The typical 
AT results have shown standard deviations of about 1 cm for 
the positions and 10 arcminutes for the attitude, respectively. 
The average boresight parameters are in Table 2. 
  
  
  
  
Offsets in IMU body Rotation [degree] 
frame [m] 
dX -1.104 o 11.160 
dY 0.047 9 2.169 
dZ 0.439 K 88.012 
  
  
  
  
  
  
Table 2. Average boresight offsets and rotations. 
6. PERFORMANCE ANALYSIS 
6.1 A performance limit of the system 
Once the boresight parameters had been established for the 
first time, a performance evaluation test was performed at the 
OSU West Campus. The objective was to test the 
performance potential of the system using operator based 
measurements. Image orientation data (EO) was provided 
from the GPS/INS navigation solution by applying the 
boresight transformation. Several models were set up and the 
control point coordinates (as check points) were measured in 
the directly oriented images and compared to the ground 
coordinates. Table 3 presents a sample comparison of these 
coordinate differences. It should be pointed out that in the 
analysis presented here, only the final positioning 
performance is addressed, which can be interpreted as the 
ultimate limit for the automated image sequence processing 
technique. 
  
  
  
  
  
  
  
  
  
  
  
  
X [m] Y [m] Z [m] 
Point 1 -0.014 0.011 0.032 
Point 2 0.015 0.045 0.060 
Point 13 0.024 -0.010 0.039 
Point 14 0.015 -0.045 -0.083 
Mean 0.010 0.000 0.012 
RMS 0.017 0.038 0.064 
  
Table 3. Checkpoint fit to ground truth. 
6.2 Realized performance 
To assess the performance of the system, absolute and 
differential tests were performed on various roads. A few 
control points were set up along a centerline at the OSU West 
Campus and were regularly surveyed from both directions. 
The acquired image sequences have been processed in two 
different ways. In monoscopic mode, the centerlines were 
extracted and the location was computed from the 
approximation that the vehicle is always parallel to the road 
surface (the image ray intersects with the base plane of the 
vehicle). To support the stereo positing, the consecutive 
images were matched to establish 3D geometry and thus 
centerlines could be determined in the 3D space. 
Experiments have revealed that the matching process is not 
only very time-consuming but it is rather unreliable as well. 
To some extent, however, the monoscopic results have turned 
out to be quite good and rather insensitive to data acquisition 
anomalies, such as gaps between lane markers or missing 
images. Therefore, we temporarily abandoned the idea of the