tor 
um 
1 a 
es, 
ent 
ent 
ne, 
Ce. 
int 
on 
sal 
  
293 
2. STEREOVISION IN AUTOMOBILE CONTEXT 
Some particular points have to be observed in an automobile environment in order to underline supplementary 
constraints to be taking in account in this context. We can summarize them in the following classes: a)geometrical 
constraints; b)time constraints; c)cost constraints. 
a) First of all, it is obvious that the objects circulating on a highway are "big", compared to the usual stereo: 
baseline (which is typically between 10cm and 25cm). Assuming B to be the baseline and F the focus lens of the 
acquisition system, we can show that the disparity d between left and right projections of an object ( size S) at a 
distance Z would be always inferior to the projection size S . The stereo geometry gives us the disparity d=BF/Z 
and s=SF/Z , so if the object size is bigger than the baseline ( S> B), then d < s (see Fig. 1). 
  
  
  
  
  
  
  
  
     
  
-—t §¢ 
«— — 
  
right retina 
  
  
Figure 1: Automobile Stereovision Constraints 
  
In other words, the two corresponding projections will be spatially overlapped, which opens the way to a local 
matching process. Besides, we can show in the same way that if the lateral distance L between two neighboring 
objects is superior to the baseline, non-corresponding projection overlaps are eliminated ( | » d), thus simplifying the 
correspondence problem, one of the main bottleneck of the approach. 
Another pertinent observation is the fact that all obstacles have vertical'edges (or almost vertical) . So we will focus 
on extracting vertical features and this is fully coherent with the fact that those features are the only one affected by 
disparity, in a horizontal stereo geometry. This remark lead us to a quasi 1-D parallel process. 
b) and c) Given the high dynamicity context, and the low expected cost of the hardware of this public 
application, we can envision a fully integrated retina-like VLSI sensor, which is a new way to deal with parallel 
processing for a hardwired real-time detector. 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommision Workshop "From Pixels to Sequences", Zurich, March 22-24, 1995