129
The simplicity of this algorithm makes it easy to implement. Besides, it does not require any data pre-processing or local
feature extraction. This facts allow to handle reasonably noise effects that occur in our range finder system. The algorithm
converges after a few iterations as showed in (Besl, 1992). Most of the computing time of the algorithm is spent in the closest
point search part (1) whose cost is O(Np*Ny), when not using optimised searching techniques. The calculation (2) and
application (3) of the transformation grow both linearly with Np.
We will now show some experiments to demonstrate the usefulness of the closest point matching algorithm for free form
object recognition. First, there will be a summary of experiments in the plane with puzzle pieces. Then, we will do a
systematic investigation of the algorithm in 3-D space using a synthetic object.
3. EXPERIMENTAL RESULTS IN 2-D SPACE
We first present experiments in the 2-D space with puzzle pieces in order to better visualise the behaviour of the algorithm in
two dimensions. The puzzle pieces are acquired with a b/w camera. After a thresholding of the image and the extraction of
ing the contour we obtain a set of points describing the border of the puzzle piece. Such sets and subsets of puzzle pieces are
s of fed into the closest point matching algorithm. Figure 1 shows some examples of the matching of puzzles and parts of puzzles
ads placed at different positions and different orientations. The example illustrates the correct matching of a complete piece and
of a subpart as well as an incorrect match of a subpart. We observe that the test pieces converge in few iterations towards
gs the model. However, the selection of the starting position becomes crucial to a successful registration.
e
1m,
atic
da
gli, test
ipe :
| i C h T AAA AR m Oo d e |
Fig. 1: Matching iterations of puzzle pieces
inly To get a global view of the influence of translation and rotation we place successively the test piece in several different
the positions around the original puzzle. Moreover the so placed test piece will be rotated around its center of mass. Figure 2
ind plots the results for all rotations and sixteen positions arranged in a grid. The black sectors at every grid position indicate the
ace angles for which the matching has been successful.
ing
ect
a [s
ble Pes
his
on.
er ® DDD
O
z eoe
en
the
= SB DD
of example
94)
the Fig. 2: Convergence zone for a puzzle piece placed and rotated differently
We observe that the translation between two pieces to be superposed is of minor influence to the convergence. However, if
the rotation angle between the test and the model piece exceeds a certain value the registration fails. The convergence zone
for the rotation angle is about between +/- 30 degrees.
Kn Since the computing time of the closest point matching algorithm grows with O(Nx*Np), we are interested in using as few
points as possible. Experiments with subsampled data confirmed that good convergence is available for reduction factors up
to eight, which allows to represent a puzzle border by only 60 points (Hügli, 1994a).
In object recognition it is important to know if the final distance between two different objects is significantly larger than
between two representations of the same object. This fact must be satisfied to allow reliable decision. Experiments with
puzzle pieces of different shapes showed that a threshold can be set for the final distance to separate reliably the object
classes (Hügli, 1994a).
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995