u————————————— 
  
296 
4. THE ANALOG IMPLEMENTATION OF OUR ALGORITHM 
The Difference of Gaussian filter is realized by an extremely precise capacitive IIR 2D spatial Gaussian filter{4], 
using only Capacitors and MOS switch. Figure 3 gives a schematic representation of such a 1-D capacitive network. 
This capacitive network is faster, more precise and smaller (in terms of VLSI-area consumption) than other 
implementations and simplifies considerably the algorithm-hardware conversion. 
ed ee ence tirer ee 
Q2 v : V v 
B: IB E av d 
Cb 
I Cr T 
Figure 3: 1-D analog Gaussian filter using switching-capacitors network 
The difference of the two Gaussians and the primitive extraction are merged together in order to avoid subtraction. 
The local extremum searching is based on charge sharing and voltage comparison and takes place only in the 
horizontal direction (in a horizontal stereo geometry only the vertical edges are affected by disparity). 
We can also show that only one static comparator is needed per node, because of the network symmetry. Indeed, 
we have to decide whenever a pixel (n) is a local extrema by comparing his value l(n) to the ones of his neighbors l(n- 
R) and I(n+R), R being the given radius. Theoretically, 2 compare operations - to the left (l(n)-I(n-R)) and to the right 
{I(n)-I(n+R)} - are needed. Let 's assume that only the left operation is done for each pixel. In this case, we can obtain 
the right operation by inverting the result of the (n+R) pixel, which computes {I(n+R)-1(n+R-R)}. 
Finally, the chosen architecture of the chip is represented in the Figure 4 below. The labeling and matching unit can 
be also implemented on host processor, while after the extraction procedure data becomes numerical. 
  
Photodiode Line-sensor 
/ Line-Scan Matrix Sensor 
  
Gaussian Filter 1 
  
Gaussian Filter 2 
  
DOG © Punu : + From the right 
0 rimitive Extraction retina 
V-I Read-Out Converter J i 
To the Host Processor 
(Labeling and Matching ) 
  
  
Line Decoder 
  
  
  
Figure 4: Schematic architecture of the retina-like detector 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommision Workshop "From Pixels to Sequences”, Zurich, March 22-24, 1995