Next:
4 Tracking the 3D
Up:
Visual Tracking of Solid
Previous:
2 Active Contour
The surface of a smooth, curved object can be described by a finite collection of parametric surface patches pieced together to form a composite surface. A single patch can be efficiently represented as (uniform) B-spline tensor product surface:
defined on a (rectangular) parametric region
, where
are the control points, and
and
are B-spline basis functions of order
k
and
l
, respectively. For basis functions of order three
is a piecewise biquadratic function with positional and first parametric
derivative continuity at the boundary between the segments (
continuity).
continuity at the border between two (biquadratic) tensor product
surface is easily achieved by interpolating the control points at the
border of the patches, choosing equal control points at the border and
assuring, that three control points across the border (with the common
border control point being the midpoint) are collinear.
Objects of genus 0 cannot be described by rectangular surface patches only. Triangular patches can be designed for inclusion by choosing identical control points at one side of a tensor product surface or by using 3-valent polynomial surfaces.
To determine the object's silhouette we compute the preimage of the
contour in the parametric regions
of the surfaces
i
first. This is a subset
of the singular points of the mappings
, where
P
is the weak perspective projection, and
is the i-th tensor product surface.
maps the points of the parametric region
i
into the image plane. For biquadratic B-spline tensor product surfaces
the singular points, characterized by a vanishing Jacobian, are the
solutions of a 5th order algebraic equation.
If the singular points have to be computed, the structure of these
curves and the visual events which might occur, have to be considered.
Presuming a smooth mapping
, then according to [
13
] the mapping is good, if the gradient of the Jacobian does not equal
zero in any point of the parametric region:
. In this case the singular points form smooth, non-intersecting curves
in the parametric region, called general folds. The curves can be
computed using a curve trace algorithm for the algebraic equation based
on Newton's method. The algorithm is initialized with the singular
points computed at the border of the parametric regions.
If the mapping
is not good in a finite number of isolated points, stating that the
gradient might vanish, a visual event of codimension 1 occurred
. The curves of singular point can only be traced from the regions
border toward those points. The existence of those points can
numerically be checked while approaching them, although their occurrence
is in practice very seldom.
The curves formed by singular points are implicitly linked across the
border of parametric regions through the curve tracing. Once these
curves are pointwise approximated, they can be mapped into the image
plane via
and are denoted by
.
The silhouette is a subset of the contour, consisting of one or more
closed curves. To compute the outer curve, the intersections and
self-intersections of
have to be computed. The outer curve
is determined and intersection points (representing T-junctions) are
explicitly marked. After a chord length parameterization,
is approximated with a closed, quadratic B-spline by minimizing the
square distance between the B-spline and the curve
. T-junctions are explicitly modeled as corners in the spline.