individual 
and the 
It systems 
  
  
points in 
ible, and 
| image 
ientation. 
by the 
| includes 
Systems. 
ing error 
lagery is 
epresents 
rameters, 
height or 
camera 
attitude 
(1) 
m, C, , 
ed “by 
erms for 
operator 
(2) 
further 
"mr (3) 
   
e 
Estimating the range of the different error terms is 
quite different for the camera exterior orientation 
than it is for the imaging system parameters. In the 
first case, the large variations in the application 
environment make it rather difficult to estimate error 
terms in general. Most of the components, such as 
the positioning data, are field-dependent. The 
imaging system, however, is quite predictable, and 
all the terms can be appropriately approximated 
after performing a calibration procedure. 
The accuracy range of the imaging system is 
determined by the resolution power of the imaging 
sensor system, the base-length, the calibration 
procedure, and the depth range of the application. 
Mobile data acquisition environments almost always 
require imaging systems based on CCD sensors. The 
image resolution of the camera used in the BNR 
project is 768 pixels by 480 pixels, resulting in 
roughly 1-2 cm nadir pixel size at a 10m object 
distance. Assuming operator positioning accuracy of 
half of a pixel, the range for the maximum 
achievable accuracy is immediately bounded at 5- 
15cm. For shorter object distances, the 5cm value 
applies, but for objects farther away, the errors soon 
become unacceptable, due to the base/depth ratio. 
The calibration of the imaging system comprises the 
determination of the interior and relative orientation 
parameters and the registration of the local camera 
coordinate system to the vehicle coordinate system, 
defined by the GPS antenna as an origin and by gyro 
determined directions as the coordinate axes. It is 
important to note that since uncalibrated cameras are 
typically used, the interior orientation should 
include a sufficient number of parameters to correct 
for lens distortion. If the calibration procedure is 
properly executed, errors introduced by the 
imperfections of parameters and random changes 
141 
Figure 3. A stereo image pair taken by GRS’s GPSVan™ (Courtesy of 
  
    
Burlington Northern) 
(such as change in the base-length due to 
temperature fluctuation) can be generally ignored 
when compared to the positioning errors caused by 
the relatively large pixel size. Our experiments show 
that using a 1K by 1K CCD sensor with 1.8m 
camera base results in positioning accuracies that 
are better than 3cm for object distances of 5m and 
15cm for 25m, respectively (He et al., 19940). 
3. RESULTS 
In 1994, General Railroad Signal Corporation 
(GRS) was awarded a contract by BNR to perform 
GPS/DR surveying and image processing services. 
GRS is the general contractor for the BNR 
engagement and is subcontracting position and 
image post-processing activities to TransMap 
Corporation, an Ohio State University spin-off 
company. BNR’s objective is to determine the 
position of tracks and the coordinates of switches 
and other wayside features at better than one meter 
accuracy. A typical stereo-pair from the BNR survey 
is shown in Figure 3 (Blaho and Toth, 1995). 
To date, GRS and TransMap have conducted field 
surveys and post-processed data for more than 
9000km along the BNR network. These numbers 
represent sufficient statistical data to analyze the 
achieved accuracy of the GPSVan™ technology in a 
real production environment. The typical BNR 
project survey lasts less than two hours and covers 
50 km. Usually there are two GPS base stations, and 
at a minimum, one quality control point (QC) is 
measured per survey. Based on thirty-five surveys of 
a 2000km railroad segment, detailed accuracy 
statistics were computed for major components of 
the GPSVan™ system. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996