SS 
aa ARS 
  
  
cena ve eus 
OE 
not required for the purpose of deriving the height of 
the robber. The error sources inherent with the simple 
Linney approach discussed above were indeed present 
but their effects were insignificant in comparison to 
the scaling error introduced by EPU. Dr Linney's 
solution was therefore originally more accurate than 
the EPU approach, despite the simplifying assumptions 
underlying it. By modifying the functional model used 
in the bundle estimation to account for differential 
scale between x and y image coordinates, a more 
accurate and precise solution could be obtained. 
2.8 Revised data processing 
Additional programming was required to allow the 
General Adjustment Program (GAP) to derive a 
differential scale factor between x and y image 
coordinates. This was incorporated into the software 
by estimating corrections to two focal lengths, one 
associated with the x image ordinates, the second with 
those in the y image direction. The differential scale 
factor was calculated from simply the ratio between 
these two estimated focal lengths. 
When these data were processed using the new version 
of GAP the scale factor was found to be 0.84 which 
represents a large difference of scale between the x 
and y image coordinates. With the effective removal of 
this particular source of systematic error it was found 
possible to include data from all four images in one 
combined solution, with two sets of inner orientation 
parameters associated with the two original cameras. 
This increased the geometric strength of the estimation 
to the extent that a parameter used to model the lens 
distortion of one camera, could be derived with 
statistical significance. 
2.0 Revised determination of height of robber 
The revised three dimensional coordinates of both the 
'foot' and 'head' points were again read into an 
Intergraph Microstation graphics design file and the 
height of robber estimated using the techniques 
discussed earlier, (Section 2.5). There was now a 
variation of 0.049 metres (1.9") between the four 
estimates, considerably lower than the range obtained 
before. Statistical analyses of the four estimations gave 
a mean of 1.781m and a standard deviation of the 
mean of 0.011m. Therefore it was concluded, with 
95% confidence, that the top of the robber’s head was 
between 1.745m and 1.817m (ie between 5° 8.7" and 
5’ 11.5") above the carpeted floor surface. 
3. CONCERNS 
The difference between the initial and revised solution 
was over 0.10m (1.888m - 1.781m) which was of 
extreme importance when put into the context of a 
    
   
   
   
    
    
    
    
   
   
   
   
    
  
   
    
    
   
   
   
   
   
   
   
   
   
   
  
  
   
    
   
    
    
   
    
    
    
    
   
   
person’s height and liberty! What concerned the 
Engineering Photogrammetry Unit most was missing 
the critical systematic error during analysis of output 
from the original self-calibrating bundle adjustments. 
The functional model selected originally was 
inadequate and examination of the output had not 
indicated this error. 
Undoubtedly a major factor in this omission was the 
poor design of the photogrammetric network, which 
was common to all three cases. If EPU had been able 
specify the number, position and type of the cameras, 
their recording rate and the number and type of control 
points the geometric strength of the network would 
have been significantly stronger. Not only would this 
have produced a precise solution but the resulting large 
and perhaps systematic pattern of residuals would have 
indicated the presence of undetected systematic errors. 
Once the correct functional model had been identified 
accurate results would then be obtained. A network 
with high internal reliability (Hottier, 1976) can yield 
both a precise, and perhaps more importantly, an 
accurate estimation. 
In these types of analyses the photogrammetry cannot 
be designed and is generally, if not always, poor. 
Although it is normally possible to obtain a solution 
this will be only accurate if the selected functional 
model accurately simulates those original ray bundles. 
In these circumstances only experience, patient 
examination and thorough testing will help the 
photogrammetrist from making potentially disastrous 
mistakes. Unfortunately the pressure of commercial 
contracts leave little time for such analysis! 
4. A CONTROLLED SIMULATION 
The application of photogrammetric techniques to 
solve this general type of problem was of interest to 
the Engineering Photogrammetry Unit and the 
concerns indicated in Section 3.0 justified further 
examination into the various factors affecting a 
solution. 
4.1 The rigid robber 
It was decided to replicate the geometric relationships 
between robber, security video camera and photo- 
control in a controlled environment. The robber was 
replaced by a survey staff located in front of an array 
of control targets normally used for camera calibration. 
The relationship between the staff and the control field 
was determined using a combination of taped distances 
and height differences. Black circular targets were 
placed over the control points and on the staff at 
heights 0.000m, 1.600m and 1.800m. These targets 
were of sufficient size to be imaged on the video