294 | 
| 
3. AN OVERVIEW OF THE ALGORITHM 
The fundamental problem in stereovision is to match the visual primitives in a stereopair of images. Marr and | 
Poggio were the first to propose a stereovision algorithm based on spatio-frequency analysis [1]. They suggested that 
this global matching problem could be reduced to a purely local one if the minimal gap between two consecutive 
primitives is greater than the maximal expected disparity. In this case, the match could be performed by a simply in-line 
search in the neighborhood delimited by the maximum given disparity. 
In order to achieve this condition, the couple of images have to be band-pass filtered, so as to eliminate high spatial 
frequency details but also their DC component. Marr & Poggio used the zero-crossings points based primitives 
extracted from the Difference of Gaussian (DoG) filtered images. In our realization we have kept this type of filter 
because of his optimal spatio-frequency properties, but we have extracted extrema points as matching primitives. 
These extrema points are observed to be more stable than zero-crossing points, and their extraction can be done by a 
| 
local comparison in the neighborhood delimited by the radius of the negative disk in the receptive field of the DoG filter 
(see Figure 2). This searching radius gives the best S/N ratio and it generates compact segmentwise responses. 
| 
Initial Impulse: 
; Primitive Extraction 
  
  
  
  
  
  
Procedure | 
DoG resr d T wh j i 
Spo i | MAX-segment 
oy Ed 
Lt ; 1 4  MIN-segment 
/ i; Radius | f 
| : | / à 
Figure 2: Extrema segment extraction from DoG filtered images 
On the other hand, if the maximal disparity value doesn't exceed half the width of the extracted segments, a direct 
local matching becomes possible between overlapped primitives. This assumes an overlap between corresponding 
segments in the stereopair but also a non-overlap between non-corresponding projections, which is assured by the 
geometrical constraint in automobile context, as discussed in the second chapter. 
The high cut-off frequency of the DoG filter must be adapted to the maximal expected disparity, given by the 
minimal distance to detect for the stereo system, and so the correctness of the matching process is insured both by the 
properly configured filter and the object size constraint. 
This stereo system has been simulated on real scenes taken from a running car on a highway and the results were 
very promising. À sequence of 16 animated images is available, giving the disparity map for a real highway dynamic 
scene. The images below show each step of the algorithm, the acquisition (Photo 1) , followed by the filtering step 
(Photo2), primitive extraction with a horizontal-labeling procedure (Photo 3), the resulting depth-map of the 
scene(Photo 4), and the depth map superposed to the initial image (Photo 5). Gray levels in the depth map represent 
  
the value of the disparity, that is white for the near object and black for the far one. 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommision Workshop "From Pixels to Sequences", Zurich, March 22-24, 1995