A KNOWLEDGE BASED SYSTEM FOR CLOSE RANGE DIGITAL PHOTOGRAMMETRY
J. Jansa and J.C. Trinder
School of Surveying, University of NSW
P.O.Box1, Kensington, 2033 Sydney, Australia
ABSTRACT:
The design of a software package for highly automated compilation of close range objects from digital images for medical examinations is presented. The
images taken by two or more CCD cameras are analysed by using techniques of pattern recognition and image matching. An automatically extendable and
improvable knowledge base and strict geometric constraints support interpretation and measurement. The goal is for the exact determination of the shape of
the object as well as the derivation of parameters describing the accuracy achieved.
KEY WORDS: Surface Model, Image Matching, CCD Camera.
1. INTRODUCTION
Despite the low resolution video images compared to images taken on
photographic film, CCD cameras are already an important tool for
mensuration in various fields, particularly in robotic systems where near
real time is necessary. The usage of CCD cameras in photogrammetric
systems where high geometric accuracy is required is limited but they are
successfully used for measuring 3D coordinates of well targeted points
(Beyer, 1991).
The project presented here in this paper deals with the derivation of a
digital surface model in particular for surgeons who need to screen
patients for monitoring medical symptoms and their correction. Real time
is not necessary but the results should be available within a few minutes.
This medical application is the first goal of our development, but ulti-
mately the development will be available for a range of applications.
Object e Knowledge
Base
LP
Qf e
88 az
go zT
m X
S
Digital | _ _ | Digital P| cC m
Image Image » m >
v Yv mo
; ; 22
Orientation -
Tool >
Figure 1: Overview
One of the main differences between inspection systems or systems for
quality control during mechanical manufacturing is the additional calcula-
tion of parameters describing the accuracy and reliability achieved.
As this tool will not be used by photogrammetric experts an appropriate
knowledge base should be included which offers support for unexperi-
enced users, reliability checks, elimination of ambiguities as well as
planning suggestions for the camera arrangements and settings (Figure 1).
2. CONSIDERATION ABOUT THE ACCURACY OF CCD CAM-
ERAS
As mentioned above the low resolution is one of the main limitations of
video cameras. Subpixel matching is the only way to obtain results of
reasonable accuracy. An important matching algorithm is the area based
least square matching. Depending on the texture and contrast of the
surface an matching accuracy of 1/3 to 1/50 of a pixel can be achieved
(Trinder et al, 1991). Tests using various patterns and contrasts showed
that for non targeted object points an average accuracy of 1/10 of a pixel
can be expected.
Independent of the methods of digital image processing, the CCD cam-
eras themselves are of poor geometric quality compared with conventional
photogrammetric metric cameras. There are many factors influencing the
geometric quality caused by the optical, mechanical and electronical parts
of the camera, the digitizing and the frame grabbing module which
decrease the internal accuracy of the image. Additional parameters
introduced into the photogrammetric bundle adjustment are able to cope
with this distortions but experiences reveal that some of these influences
are unstable and vary even during data acquisition. A calibration of the
camera under laboratory conditions is not sufficient as it can only give an
overview of the behaviour of the elements of the acquisition system and
the various settings.
Our system uses a bundle adjustment with the following orientation
parameters:
Interior orientation:
- Focal length (not adjusted)
- Separate image scales for x and y
- x,y coordinates of the principal point
- cubic and power 5 term of lens distortion
- linear skewing factor for scan line shift
Exterior orientation:
- X,Y,Z coordinates of projection centre
- three rotation angles
It is absolutely necessary to repeat the calibration in situ before data
capture and if possible to include terms for selfcalibration in the math-
ematical model used for compiling the image data.
Preliminary tests showed that distortions caused by our digital system
(Minitron CCD Cameras with SONY 16 mm wide angle lenses, ITEX
Monochrome framegrabber) can be modelled with the bundle adjustment
mentioned above. The power of 5 term of the lens distortion and the
skewing factor were of least significance and only a high accuracy
calibration might return reliable values for these parameters.
For handling all data of the bundle adjustment and for a quick editing
possibility all observations and adjusted data are organized in tables.
(Figure 2). There is a table for the object space which contains the object
coordinates of all control points, a constraint table which contains mainly
the observed distances in object space and an image space table which
includes the coordinate lists of all measured image points and the provi-
sional and adjusted values of the orientation parameters. The data of all
tables can be edited.
