New frame 
Reset Start processing End processing 
FIFO A 
data 
Field 1 Field 2 
  
l| oda| ' | | | | 
  
  
  
  
  
  
  
  
  
  
|| field | | | 
| Tim 
Le k CA A | > 
| | | | 
FIFO B' | | | 
data | | | | | I | 
| | | 
| ; | : | : | 
i| Even 
| | field | | | | 
| oy po Time 
| i | 
| | | 
Image 1 Image 2 Image 3 
Figure 2 Graphical representation of filling the FIFO's with 
data and frame alignment after reset. 
2.2 Image processing 
2.2.1. Target image processing. Image acquisition is a 
fundamental process for photogrammetry. Prior to obtaining 
any video data, the VFE-90 module must be reset. For an 
interlaced imaginary a new frame and subsequent frames must 
be aligned. Then, the video signals are processed line by line. 
Three types of data: sub-pixel location, intensity, or new frame, 
are read out from the FIFO's. Decoding the data gives 
meaningful edge and intensity information of target sections. 
Figure 3 illustrates extracted data from the FIFO's with a 
threshold set at 50 for three targets (two of which overlap each 
other in the same line) in an interline image. The threshold 
value is set by an 8-bit I/O port on the GPIO-90 board. The 
lower two bits of the intensity are not used so a threshold of 50 
equates to 200. At this stage in the development of the 3D-NET 
system the video signals have not yet been adjusted optimally. 
Line ! coln 
Line_2 eoln 
Line 377 eoln 
Line 378 eoln 
Line 379 edge 354 267 302 286 271 252 205 eoln 
Line 380 edge 353 244 335 385 385 349 305 243 coln 
Line_381 edge 353 249 339 395 397 355 315 248 eoln 
Line 382 edge 353 214 279 317 306 280 260 209 eoln 
Line_383 edge 355216 211 eoln 
Line_384 coln 
Line_385 eoln 
Line 386 eoln 
Line 387 edge 346 272 331 354 326 295 256 coln 
Line 388 edge 346 219 254 261 246 229 206 coln 
Line 389 edge 346 254 300 309 294 264 227 edge 356 254 278 265 245 226 eoln 
Line 390 edge 345 209 295 358 370 349 308 256 edge 355 250 332 366 348 324 282 214 coln 
Line 391 edge 347 210 212 204 edge 355 246 325 359 342 317 275 210 eoln 
Line 392 edge 355221 269 287 279 255 228 coln 
Line 393 edge 357218 213 210 coln 
Line 394 eoln 
Line 581 coln 
Line_582 coln 
eoln = end of line marker 
edge = beginning of a new line object 
Line number are calculated by counting eoln markers 
Figure 3. Data Extracted from FIFO’s A and B 
The target location algorithm only requires edge data belonging 
to two consecutive lines of an image at a time to compute the 
sub-pixel location of each object. À parameter buffer is used to 
store the peak intensity, a summation of intensity and 
summation of x or y location times intensity for each object. 
The data for consecutive lines are stored in two buffers “ping” 
and “pong”. The ping buffer of current line becomes the pong 
buffer for next line target recognition. The edge pairs in the 
ping buffer are compared with those in the pong buffer to 
ascertain the state of the edge pairs in the current and previous 
lines. By comparing the starting and finishing pixel values of 
target sections in each two consecutive lines, the targets of 
legitimate shapes present in the frame are reconstructed. 
Splitting or merging targets are recognised and flagged as 
invalid photogrammetric target images. The sections of invalid 
targets within subsequent lines are processed but no sub-pixel 
location is computed. The completely reconstructed targets are 
assessed for validity using the area and peak parameters. 
2.2.2. DSP programming issues. The ADSP-21xx family base 
architecture ^ provides  single-cycle = computation for 
multiplication together with accumulation and supports 
extended sums-of-products. Detecting target edges and 
accumulating grey scale intensity value for each target can be 
achieved efficiently with a DSP. These data are used to 
calculate the grey-scale centroid of target. 
2.3 Image location 
2.3.1. On-the-fly centroid computation. The main task of the 
image location algorithm is to calculate the grey scale centroid 
of the targets. Object of all shapes have to be processed but 
only the objects that meet the criteria set by the target image 
recognition algorithm have to be located accurately. One 
method of computing the centroid would be to create sub- 
images (together with an offset from the origin) for 
conventional processing. This would involve storage of the 
image in a temporary location and extra computations. The DSP 
has limited resources for such tasks and therefore another 
method was chosen. This method uses the line-by-line approach 
used by the image recognition algorithm to compute the 
centroids at the same time as the objects are being recognised. 
In this way the storage of information is limited to the data for 
the summations required for each of the objects. The DSP is 
highly efficient at multiply and accumulate operations required 
in the centroid calculation. However, the DSP used is a fixed- 
point device which multiplier produces a 32-bit product. When 
the accumulation has overflowed beyond the 32-bit boundary 
the object is unlikely to be a retro-reflective targets and so the 
accumulation stops, but the algorithm must still deal with the 
object to avoid problems in subsequent lines of the image. 
2.3.2. Criteria for selection of legitimate targets. The 3D- 
NET system was designed with the use of retro-reflective 
targets in mind, these targets produce images of predictable size 
and shape. Other features above the threshold will often be of a 
different size and shape. To recognise target-like features one or 
two measures can be used on-line to select candidate targets. 
Two intuitive and easy to implement measures, which 
nevertheless give most of the information required, are the area 
and peak intensity. The area is simple to accumulate on-line 
being the sum of line lengths for each object, the peak requires 
the comparison of current intensity with the previous intensity 
and the storage of the greatest. Other measures such as radius or 
  
  
maximum diffe 
are likely to prc 
2.4 Results 
A series of test 
the intelligent c 
on the image d. 
the image size 
targets were sat 
The first test ev 
an overflow of 
processing can 
computed for v 
1. 
Avera Al 
Table 1. The 
proce 
In a typical sce 
be oriented of 
relatively dim 
orientations. A: 
30 pixels wouk 
to 340 pixels 
figure is indep 
function of sto 
encoding schen 
can be stored 
represents a « 
threshold proce 
be the same reg 
used if the targ: 
A second serie: 
capability of tl 
processed to as 
a FIFO overflc 
DSP is incapal 
make this test | 
also used so the 
and peak of ea 
results. 
Area of tar 
2 
É 
E 
E 
Table 2. Real 
The results sho 
some 6 - 7 pix« 
target images «