Biomedical Engineering Reference
In-Depth Information
Matching is performed using the contour codes of points on the two surfaces.
Fast indexing was achieved by hashing the codes for all models in the library
into an index table.
At runtime recognition, the splashes of highly structured regions are com-
puted and encoded using the same encoding scheme. Models which contain
similar codes as the splashes appearing in the scene are extracted. Verification
is then performed for each combination of three correspondences.
1.5.1.2
The “Point Signature” Surface Registration
This approach was introduced by Chua and Jarvis [28] for fast recognition.
They establish a “signature” for each of the given 3D data points rather than
just depending on the 3D coordinates of the point. This is similar to the “splash”
representation but instead of using the relationship between the surface normals
of the center points and its surrounding neighbors, they used the point set itself.
For a given point
p
, they place a sphere of radius r, centered at
p
(simi-
lar to the method used in splash as depicted in Fig. 1.6). The intersection of
the sphere with the object surface is a 3D space curve whose orientation can
be defined by a directional frame formed of the normal to the plane fitted to the
curve, a reference direction and their cross product. The next step is to sample
the points on this curve starting from the reference direction. For each sampled
point there exist two information, the distance from itself to the fitted plane and
the clockwise angle about the normal from the reference direction. Figure 1.7
shows some typical examples of point signatures for different surface types.
Due to this simple representation, the 3D surface matching is transformed
into 1D signature matching. In their paper they analyze this signature matching
and estimate the accepted error tolerance in the matched signature. Prior to
recognition, the model library is built by computing the point signatures for
each point in the model and for every model. They also used hashing to index
the signatures in a table. At runtime, the surface under study is sampled at a
fixed interval and the sampled points are used in the matching process.
1.5.1.3
The “COSMOS” Surface Registration
The goals of the COSMOS (Curvedness-Orientation-Shape Map On Sphere) rep-
resentation scheme were, first, to be a general representation scheme that can
