Figure 5. The final results of target recognition. 
A number is given to each of the targets in the order in 
whichit is recognised and a cross placed at that position. 
The use of a normalised images and binary 
segmentation allows for efficient recognition of the 
targets. While other methods using a grey scale image 
may be contemplated most will be more 
computationally expensive. 
3. TARGET LOCATION IN A GREY SCALE 
IMAGE. 
Grey scale proces ing uses the image intensity values at 
each pixel site directly. In this case the image blurrin 
that inevitably occurs in the imaging process can be use 
to match not only shape but expected intensit 
distributions. Cross sections of a small round target will 
have a Gaussian shaped intensity distribution. It is 
possible to erm image matching in grey scale but 
this method will be computationally more expensive. 
Hence, the grey scale image is only used to provide the 
precision location of the target to subpixel accuracy. 
Once the legitimate targets have been identified it is 
then necessary to refine the precision with which they 
are located in the image. A survey (West & Clarke, 
1990) of subpixel techniques using grey scale images 
indicates that high location accuracies have been 
pou by a number of authors for objects such as 
edges, Gaussian blobs, patterns, etc. À number have 
demonstrated that the centre of the targets used in 
photogrammetry are no exceptions (Trinder, 1989; 
Wong, 1986). The subpixel accuracy with which a target 
can be located using these methods has been reported 
as high as 0.01 of a pixel. Accuracy is limited to a large 
extent by noise (Deng,1987) if linejitter (Beyer, 1990) is 
not a problem. 
Many methods have been used for subpixel location of 
targets, among them are interpolation, correlation, 
centroiding, differential, and shape fitting, In this study 
the centroiding method was chosen for investigation 
because the computation will give consistent results on 
small circular shaped targets even when viewed from 
differing angles. 
The initial positions of the targets given by the binary 
image are used to place a small rectangular window 
around the target. A suitable threshold was used to 
eliminate the background from the target. A 15x15 pixel 
window was found to be a suitable size. From this 
window the precise centre of the target was calculated 
using the well known Equation 6. 
nm 
x = MZ Xj gij 
i=1j=1 
nm 
y 7 yMZ Zi gi (6) 
i=1j=1 
Where 
nm 
M-2ZZgi 
i=1j=1 
gij is the grey scale value of each pixel and n = m = 15. 
The results of the target coordinates for each image 
were then stored in a file for subsequent labelling with 
respect to the reference image. Following the labelling, 
described in the next section, the ce adjustment 
procedure (Granshaw, 1980) was used to calculate the 
coordinates of the targets. The results gave a global 
variance factor for the procedure of 1.578 which could 
be equated to a subpixel accuracy of 0.13 of a pixel. 
A problem has been reported using the centroid 
method (Trinder, 1989) because of the influence of the 
background. This will not affect targets which have a 
uniform background illumination, but will cause some 
asymmetry in the image of targets with a nonuniform 
background. This effect is illustrated by viewing a small 
section of the profile of the image shown in Figure 1. 
Although this is just a single slice of the target image it 
demonstrates the problem. 
256 T T i T 
224 
192 - 
160 
128 } FA - 
96 nil x 
64 + « 
32 
0 1 I l 1 
40 45 50 55 60 65 70 75 80 85 90 
  
Intensity / bits 
  
  
  
Pixel position / n 
Figure 6. Section of target image shown in Figure 1.