ABSTRACT:
We present an algorithm to detect and recognize road signs from embedded terrestrial images. We argue that the most important
process in such algorithms is the fine detection which isolates object shape in images. After this step, recognition can be processed by
simple normalized template matching. We then use a hybrid evolutionary algorithm capable of performing the fusion between colour
and edges detection. This hybrid approach associates a stochastic process and a local deterministic error minimization to increase the
precision and improve the repeatability of the convergence by eliminating certain unpredictable processes such as mutation. Primitive
fusion brings precision and decreases the necessary number of iterations (about 5 times faster) required to optimize the influence of
every primitive during the algorithm execution. We present this algorithm and show that we can use a final template matching in a
simple way.
1 INTRODUCTION
Over several decades, digital technology has become a crucial
tool and data ownership has become the key to knowledge. Dig
ital data allow us to understand our environment and optimize its
management whatever the application. Digitalization tools have
thus become a real need.
Road network data is a particularly pertinent example since such
networks represent huge responsibilities for their administrators
when roads contain defects and incoherence which can provoke
accidents. There is therefore a growing trend to the digitalization
of knowledge of road networks and their equipment to improve
management, security and, in the near future, lead towards au
tonomous navigation in urban road environments using advanced
maps.
Consequently, laboratories and companies are interested in de
veloping mobile systems to acquire digital descriptions of roads.
These systems, called mobile mapping systems (MMS), produce
huge quantities of data based on multiple sensors such as cameras
and lidars. These data are necessary but. nevertheless, insufficient
because they do not bring semantic knowledge about the environ
ment. Acquisitions are still limited to small areas to prevent non-
exploitable quantities of data being acquired from MMS. For this
reason the current aim of many researchers is to develop robust
algorithms to automatically recognize objects by image or signal
processing, and to limit the manual treatment by human opera
tors.
Algorithms of shape recognition by image processing can be gen
erally separated in two distinct tasks : detection and recognition.
We think that the most important task is detection which consists
in producing smart samples of the image data. Afterward, these
samples can be compared to a database of known object images.
We therefore developed an algorithm able to determine the affine
mapping between the image of the object and real spatial config
uration of the object. In this way we are able to obtain a quasi-
perfect front view of the object in a sample image. The sample
image is then used in a simple Normalized-Grayscale correlation
to recognize the correct object.
To do this, we oriented our research towards hybrid evolutionary
strategies (Hybrid-ES). Evolutionary Algorithms (EA) are sto
chastic processes for optimization problems regrouping genetic
algorithms (GA) and ES : biological metaheuristics simulating
natural phenomena such mutation and natural selection, keeping
precise scheme. On the other hand, some algorithms use deter
ministic methods to match two different shapes like, for example,
the iterative algorithm 1CP for Iterated Closest Point introduced
by (Besl and McKay, 1992).
We propose to combine both approaches and to fusion edge and
colour extraction to improve precision and convergence speed. In
what follows we start with a rapid overview of related work deal
ing with road sign recognition and general shape recognition, we
then recall some previous knowledge about deformable templates
and finally we present our algorithms, results and conclusion.
2 RELATED WORKS
Road sign recognition is a vast topic in image processing. Ini
tially, accumulative methods such (Hu et ah, 1998) and (Peder
sen. 2007) seem interesting to detect manufactured shape in im
ages. These methods permit the extraction of polygon centres,
lines or circles thanks to statistical processes. Some of them used
these accumulative methods to the particular application of road
sign recognition application such as (Belaroussi and Tarel, 2009)
and (Barnes et ah, 2008). The complexity of the search increases
with the complexity of the polygon we are looking for, and in
the end. the accumulation space becomes very difficult to sam
ple. Besides, these methods show low tolerance to strong affine
transformations in the image (circle easily becomes ellipse for
example).
On the other hand, (Arlicot et ah, 2009) proposes to pre-detect
colour areas and to filter these connected components to find el
lipse equation using a RanSaC algorithm. This algorithm is only
available for circular road sign also it seems to be robust to spatial
transformation.
The seminal works of (Jolly et ah, 1996) use a Simulated Anneal
ing algorithm (SA) using deformable templates to detect vehicle’s
profile and use motion detector to help the convergence of the
system. Extending this approach based on deformable template.
M.Mignotte proposes three kinds of optimization algorithm to de
tect particular shadows in sonar images : simulated annealing,
287
