Figure 4: Edge extraction: red crosses represent subpixel 
accuracy edge position. : a) extracted edges, b) a zoom on 
edges of (a), 5 points are chosen for tangent estimation, c) 
Difference between pixel accuracy tangent and sub pixel 
one. 
(a) (b) 
Figure 5: (a) Example of all the centers and axes explored 
by RANSAC algorithm (b) the estimated solution. 
Figure 8 shows some result of correlation. We match de 
tected red signs only with the red reference signs and blue 
ones with blue references. However in Figure 8, corre 
lation coefficient with all signs are shown to demonstrate 
the discrimination power of correlation function. In most 
of the cases, the maximum of correlation coefficient corre 
sponds to the good sign. We accept the maximum of corre 
lation if it is higher than 60%. Hypotheses with lower cor 
relation coefficients are rejected. This threshold is chosen 
relatively low. The reason is that the texture of signs in im 
ages suffer from both radiometric calibration problem and 
illumination changes within one sign. Better radiometric 
calibration can partially reduce this effect. So higher cor- 
Figure 6: (a), (c) and (e) are the original image windows 
and (b), (d) and (f) are respectively their resampled images. 
09 0® 
(a) (b) (c) (d) 
0@© 
(e) (f) (g) (h) 
0 90® 
(i) 
(m) 
(j) 
(k) 
(o) 
(1) 
Q @ © 0 
(n) 
(p) 
Figure 7: Circular road signs reference database. 
relation coefficient thresholds can be set in the algorithm 
and improve the reliability of recognition. 
Score corr (A, B) = 
E E [A{x,y)-A\[B(x,y)-B] 
x=ly=1 
E E [A^X^-A]* jr £ [B(x,y)-B] 2 
c=ly=l x=ly=l 
(4)