The software of the system consists of three main 
parts: 
Recording 
Tracking 
Analysis 
2.1 Recording 
The recording transfers a sequence if images from an 
external source to the internal Video Disc of the 
TrackEye system. A video disc file can be generated 
using three different sources: 
Standard video sources from a video or a video 
tape recorder. 
Film scanner for 16 mm cine film 
Selected copies or sequences from earlier 
recorded video disc files. 
22 Tracking 
The tracking is automatic with manual supervision 
and is able to follow 95 objects simultaneously. The 
points are defined by the user in the first image by 
positioning a cursor on the point of interest. 
Interactive tools for zooming, backtracking, motion 
view survey etc. are provided. The manual 
supervision can be used in several ways depending 
on the complexity of the sequence, from manual 
marking with position proposal to fully automatic 
tracking with an error list indicating the uncertain 
areas in the sequence. The tracking algorithm uses a 
modified correlation technique and 
23 Analysis 
Point positions from the tracking session are used to 
perform calculations of distance, velocity, 
acceleration, angles and angular velocities. Other 
features are: 
Result presentation with graphical overlay on 
the image sequence 
Data and result filtering 
Comparison with previous evaluations 
Image and result data base 
Report and documentation generation 
The new module for the 3D calculations were to be a 
part of the analysis software package. Since an 
existing data structure was present, the module had 
to fit into that environment. It turned out to be 
rather complicated to do this, and the time schedule 
for that part of the work were heavily under- 
estimated. 
3. CHOOSING AN APPROACH 
The main choice was to decide which type of basic 
mathematical model to use, either projective 
geometry, leading to the direct linear solution, DLT, 
or the collinearity equations, leading to the bundle 
adjustment solution. The given conditions can be 
describred as: 
Accuracy requirements: + 10 mm 
2 - 6 high speed cine film cameras 
No fixed camera positions between sequences 
No fiducial marks in camera 
No fixed focus cameras 
Zoom lenses may be used 
Unstable inner orientation during a sequence 
Several of the above mentioned conditions were 
pointing toward the DLT solution, but the need of 
corrections for radial distorsion and scale differences 
in x and y and the fact that the inner orientation 
might change during a sequence made the bundle 
adjustment more attractive. If a change were detected 
during the sequence, a correction of the inner and 
outer orientation parameters would require less 
control points to compute with a self-calibration in 
the bundle adjustment than using the DLT. As a 
consequence of this, the bundle adjustment approach 
was chosen, giving a good structure for a self- 
diagnosting system with as little operator support as 
possible. The GENTRI (Larsson) soft-ware was used 
for the bundle adjustment and self-calibration. 
One of the major disadvantages of the chosen 
appoach is the need of approximate values for the 
camera orientations. 
For a comprehensive comparison between the DLT 
and bundle approach see (Edgardh, 1992) 
4. IMPLEMENTATION 
4.1 Overall view 
The complete system for the 3D measuring can be 
divided into two main parts: point measurements 
and calculations. The point measurements is done 
automatically by a tracing algorithm from one image 
to the next. This part of the system will not be 
discussed here but is well documented in other 
sources (ref ). The mathematical (or photo- 
grammetrical) part must interact with the point 
measurements with a high flexibility to be able to 
meet the level of automation. The three main parts 
of the calculations, calibration, bundle adjustment 
with self calibration and 3D measurements, must for 
this reason be able to adjust themselves to the needs 
at the different stages in the procedure. 
Normally a sequence of images, taken with the speed 
1000 frames/seconds, is calculated sequentially and 
the information regarding orientation and calibration 
is brought along and used if relevant. As can be seen 
in fig 4 the different actions taken by the system will 
be guided by the stability and functionality of the 
calibrated camera parameters. If the inner orientation 
parameters are stable, then the calculations can be 
done with a prior calibration of a known test-field 
and no additional self calibration is necessary. If the