Road Sign Detection and Recognition
Using Perceptual Grouping
Stéphan Prince, M.Sc. student
Robert Bergevin, advisor
This project presents a new method for detecting and recognizing road
signs on the basis of their shape. The detection step is based on a geometrical
analysis of the edges and groups extracted from an intensity image of the
sign obtained from an arbitrary viewpoint. It is able to identify triangular-shaped
and rectangular-shaped road signs under different viewing conditions. After,
the recognition step validates and classifies the detected road sign on
the basis of the shape of its inner symbols to improve the reliability
and the robustness of the system. The analysis and modeling phases are
based on a step of preprocessing, where each road sign model is described
by a minimum set of interest points invariant in space. These interest
points are obtained by (i) approximating the outer and inner road sign
contours using straight line segments and circular arcs and (ii) extracting
all junctions points between pairs of lines, arcs and line-arc. Different
images have been successfully tested with this method.
Introduction
The ability to recognize objects in cluttered scenes
is essential to the functionality of a partly on totally autonomous mobile
system. In most techniques proposed in literature, the road signs present
no occlusion. Also, no such method is able to verify recognition hypotheses
under the various possible imaging conditions [1] [2]. For this purpose,
the proposed approach is based on a correspondence method known as "Geometric
hashing" [3]. Under this approach, several matching features are possible:
points, lines and curves. In the present work, we developed a system based
on point and line matching techniques to recognize road signs under arbitrary
views from intensity images.
Road signs representation
Before recognition operations, a preprocessing of
one or several models is executed. During this step, the information of
each model is indexed in the reference table using a minimal number of
relative features which are (i) invariants with respect to the transforms
(rotation, translation and scaling) between model and scene and (ii)
which make it possible to deal with occlusion. This modelling phase
is realized using the same algorithm as the one used when processing a
target image. An image acquisition of road signs have been realized in
laboratory with an uniform background. The obtained models are made up
of straight line segments and circular arcs. To facilitate the computation,
a circular arc is always modelled as two straight line segments connecting
the first, middle and endpoint of the arc. With this approach, the processing
is uniform for any road signs type.
Road signs detection
We exploit organizational principles to extract
the 2D structural information of the road sign contours. The perceptual
grouping extraction is realized using five processing steps:
a) preprocessing of an intensity image to extract
the edges of the frame and inners symbols,
b) linking process to build connected contours,
c) segmentation of these contours into straight
line segments and circular arcs,
d) extraction of groups of connected segments between
any pair of straight line segments and circular arcs,
e) identification of closed contours.
With closed contours and groups of junctions between
straight line segments and circular arcs, it is possible to detect road
signs in complex scene.
Interest points considered include all junction
points connecting lines and arc segments on closed contours
Recognition
During the recognition phase, a voting procedure
is executed using the model reference table to determine the correspondence
between three interest feature points of the model and those in the segmented
intensity image of the scene. For this project, the model objects and the
scene are described by sets of interest points, which represent the corners,
points of sharp concavities and convexities.
The notion of affine coordinates is used to compute the
invariant of a point V given three non-collinear points (A,B,C). The point
V is expressed as:
The affine transform approach may allow recognition of
occluded objects since the point coordinates of an occluded object in the
scene will have a partial overlap with coordinates of the stored model.
This is true if both objects are represented using the same triplet of
points as basis. However, this dependence of the representation on a specific
basis triple may preclude recognition when at least one of the basis points
is occluded. Hence, we represent the object points by their coordinates
in all possible affine basis triplets in order to enable recognition of
the objects with partial occlusion.
Linking interest points
This step connects all extracted interest points
on a contour to form a closed polygon. This polygon is extracted to make
possible the verification of the recognition. Two interest points will
be linked if both are on the same segment. Also, a post-processing step
was added to find missing junctions in order to complete a closed polygon.
Determination of the pose
The affine transform is also used for the determination
of the pose, that is the position and the orientation of the road sign
with respect to the camera. The precision of the computed pose depends
on the image quality (model or scene). However, the pose has been computed
with success for different road signs.
Acknowledgments
This project is realized in collaboration with Centre
de Recherche Informatique de Montréal (CRIM).
References
[1] G. Piccioli, E. DeMicheli and M. Campani, "A
Robust Method for Road Sign Detection and Recognition", Lecture Notes
in Computer Science, Vol. 800, 1994.
[2] M. Betke and N.C. Makris, "Fast Object Recognition
in Noisy Images Using Simulated Annealing", Proc. of ICCV, 1995.
[3] Y. Lamdam and H.J. Wolfson, "Geometric Hashing:
a General and efficient Model-Based Recognition Scheme", Proc. of ICCV,
Dec. 1988.
[4] J.L. Arseneault, R. Bergevin and D. Laurendeau,
"Grouping of Straight Line Segments and Circular Arcs for Scene Analysis",
Vision Interface 94, Banff Alberta, May 1994.
|