). In this
undaries
Canny's
id apply
to the
'ctances,
1 for the
n easily
'ojection
a coarse
asier to
ll errors
mination
obtained
generate
ints) by
f central
sections
rmine a
raced to
to work
ures, by
ive us a
“the 3D
ondence
is based
sed onto
| in pairs
images.
linearity
occluded
uple the
re of 3D
on the
long the
d out to
s which
omputer
Ber97]).
because
graphies
ws. The
/deletion
mentary
to the
a simple
raph Gr
nd their
riangles.
folds the
ith the
ation of
y are no
je image
cclusion
ed than
in above
ts along
jents are
inserted/deleted by modifying the corresponding graph
associated by using adjacency relations in the same way as for
the Delaunay graph. Thus, the insertion of a new segment
generates at most three new regions which are represented in the
updated graph as a ternary unfolding of the original graph.
3. Trapezoidal maps for grouping
We have focused towards an automatic generation of robust
grouping criteria onto each image, instead of using more
traditional comparisons between successive images with
epipolar constraints. Grouping criteria are represented by
bilinear incidence conditions linked to deformed rectangles with
an apparent motion. Hence, our emphasis is put directly onto
shear transformations involving the projection of rectangular
regions in the world. Redundant information linked to the
estimation of trapezoids supporting geometric transforms is
robust. Furthermore, it allows us to maintain incidence and
adjacent constrains with an easy discrimination between interior
and exterior regions w.r.t. reference lines.
To avoid an excessive number of triangles with abrupt changes
in the depth, we have adapted the trapezoidization algorithm
described in [Ber97] by including degenerate cases. Main
restrictions for the selected data concern to the length and
orientation of segments. Such restrictions allow us to simplify
the treatment of meaningful information as much as possible. In
this way, we obtain less regions than for the original algorithm
described in [Ber97]. This lower number of candidates
simplifies the local grouping and the global generation of coarse
perspective models.
The insertion/deletion of segments plays the role of elementary
events, and it allow us to update the available information
associated to a trapezoidal graph. The symbolic representation
associated to such a graph plays the same role as the Delaunay
graph linked to a triangulation. To accelerate the comparison
between trapezoidal maps associated to a sequence of views, we
reduce ourselves to the study of complete trapezoids inside a
vanishing cone Cy associated to the projection lines with vertex
the vanishing point py. Such cone is symmetric w.r.t. to the
horizontal vanishing line /oo
The lack of symmetry w.r.t. the perpendicular to the vanishing
line through the vanishing point give us information about the
relative orientation of the platform w.r.t.
the lateral walls. The relative orientation is evaluated in terms of
the ratios between homologue trapezoidal regions located at left
and right sides of the perpendicular. Let us see how to manage
such trapezoidal regions.
We construct 7rapezoidal maps as a mid-level grouping after
generating a coarse perspective model for the scene without
reconstruction. Let us see some details. We decompose our
construction in the following steps: initialization, tracking of
homologue elements, propagation, prediction and errors
estimation. The main novelty concerns to the homologue
elements which are not points, not even segments, but
trapezoidal regions. In our case, the reduction to the 1-
dimensional case is possible thanks to trapezoids have always
two vertical segments. Hence, it suffices to evaluate how
trapezoids are stretching, decomposing or grouping following a
list of easily updatable events. Let us see how to initiate the
process.
Canny's minisegments are grouped according to collinearity
constraints giving a collection of vertical large segments. Skew
projection lines are obtained by applying a regression technique
to the extremes of large segments. Next, we compute the
pairwise intersections of projection lines. By using a weighted
average for such six intersections around the vanishing line, we
determine the vanishing points and retrace some "virtual"
projection lines starting from the computed vanishing points.
The intersection of such virtual projection lines with the upper
and lower boundaries of the image determine a light cone C.
We construct a trapezoidal map inside the cone of light (by
adapting [Ber97]). Elementary events are given by meaningful
segments. The trapezoidal map is easily updated in an
incremental way by inserting/deleting trapezoids associated to
elementary events given by segments. The splitting/grouping
process arising from such updating can be described by an
algorithm with a linear complexity in the number of elementary
events.
Some references
[Agu01] S.Aguilar and M.J.Antolínez: "Estimación del
movimiento aparente a partir del análisis de escenas
estructuradas de interior", Master's thesis, Univ. of Valladolid,
2001.
[Avi00] S.Avidan and A.Shashua: "Trajectory triangulation:
3D reconstruction of moving points from a monocular image
sequence", IEEE Transactions on Pattern Analysis and Machine
Intelligence, 22(4), 348-357, 2000.
[Ber97] M.de Berg, M.Van Kreveld, M.Overmars, and O.
Schwarzkopf: "Computational Geometry. Algorihtms | and
Applications", Springer-Verlag, 1997.
[Coe92] C.Coelho, A.Heller, J.L.Mundy, D.A.Forsyth and
A.Zisserman: "An Experimental Evalaution of Projective
Invariants", in J.L.Mundy and A.Zisserman (eds): "Geometric
Invariance in Computer Vision", The MIT Press, 87-104, 1992.
[Fau93] O.Faugeras: "Three-dimensional computer vision", The
MIT Press, Cambridge, MA, 1993.
[Har00] R.Hartley and A.Zisserman: "Multiple View
Geometry", Cambridge Univ. Press, 2000.
[Hor81] B.K.P.Horn and B.G.Schunk: "Determining Optic
Flow", Artificial Intelligence 17, 185-203, 1981.
[Ma98] Y.Ma, J.Kosecka and S.Sastry: "Motion Recovery from
Image Sequences: Discrete Viewpoint vs. Differential
Viewpoint", Computer Vision, ECCV'98, 5th European
Conference on Computer vision, Vol.II, LNCS 1407, 337-353,
1998.
[SGB98] M.C.Shing, D.B.Goldgof and K.W.Bowyer: "An
objective methodology of edge detection algorithms using a
structure from motion task", Proc. of the IEEE Intl Conference
CVPR, 1998.
[Tia96] T.Y.Tian, C.Tommasi and D.J.Heeger: "Comparison of
Approaches to Egomotion Computation", Computer Vision and
Pattern Recognition, IEEE, 315-320, 1996
[Zha92] Z.Zhang and O.Faugeras: Springer-Verlag, 1992.
—151—