In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C., Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part 3A - Saint-Mandé, France. September 1-3. 2010 
Figure 11 : scores for mid points candidates 
The output points will be further fitted by a Non Uniform Rational 
Bezier Spline (called NURBS hereafter). Using scores as weights 
in NURBS computations seemed at first a good idea, but as even 
a small weight influences the final curve, the threshold approach 
has been preferred. 
3.4 NURBS Fitting 
Due to the algorithm architecture, road boundary candidates are 
generally quite sparse and irregularly spaced, so a simple lin 
ear interpolation between them would not be satisfying. Resul 
tant points are thus approximated by 3D NURBS, parametric 
curves whose resolution is directly related to the distances be 
tween successive knots (Cf. figure 12(a)). Applying algorithm 
described in (Peterson, 1992), we output 3D curves w-ith a con 
stant arc-length spacing (Cf. figure 12(b)). This method is based 
on spline arc-length estimation, using Gauss-Legendre quadra 
ture and Newton's root finding method. 
Figure 12: after reparametrization, curve points are equally 
spaced 
4 APPLICATIONS 
4.1 Road curvature estimation 
Using the NURBS computed in section 3.4, it is almost straight 
forward to determine road curvature with any desired resolution. 
We chose a 50 cm resolution on the central NURBS, and com 
puted the circle from three successive points. Two results can be 
seen on figures 13(a) and 13(b), where point color changes from 
green to red when the curvature increases. Hazardous areas can 
thus be detected and precisely reported on a geographic informa 
tion system. 
4.2 Road width estimation 
In the same manner as in section 4.1, we chose a 50 cm reso 
lution on the central NURBS, and computed at each point the 
curve normal plane, reminding that we deal w ith 3D curves. It is 
then to find the intersections betw een this plane and the left and 
Figure 13: road curvature estimation for two different areas 
right splines to obtain the left and right boundaries. Following the 
same path than in section 3.4, we applied the Newton's method, in 
order to minimize the distance between the plane and the bound 
ary splines. The minimization starting point is here crucial and 
has to be carefully computed. 
4.3 Road signs extraction 
Road signs are designed to present a high reflectivity ; on im 
age 6, where the points reflectivity is displayed as a gray level, 
we can observe white areas, corresponding to license plates or 
road signs. Using a simple threshold on reflectivity value then on 
the size of the resulted areas, we can easily extract road signs in 
the virtualized environments (Cf. figure 15). The sign areas can 
further be reprojected on image in order to perform optical fine 
extraction and recognition. 
?■<'?// ! I 
Figure 15: road signs are displayed in blue 
4.4 Road markings extraction 
With a correctly extracted road and using the reflectivity data, we 
can also extract road markings trough a simple thresholding. As 
claimed (Dietmayer et al., 2006), asphalt presents a much lower 
reflectivity than road markings so that threshold determination 
is quite easy. Nevertheless, this approach can in some cases be 
less robust than image processing methods, as it highly depends 
on marking reflectivity, which is faster deteriorated than white 
painting. 
5 CONCLUSION AND FUTURE WORK 
We presented here a new acquisition platform dedicated to road 
environment reconstruction, surveying and interpretation. Real- 
285