Image Processing Reference

In-Depth Information

1 Introduction

Registration is a process to find correspondence between data sets, and generally could be di-

applications input data set are represented as point cloud (segmented surfaces of the objects,

etc.). Then in the processing pipeline data sets should be register, to find the correspondence

between data sets. The most popular approach, in this case, is iterative closest point algorithm

(ICP) [
2
]. This algorithm was proposed by a few researchers independently [
3
,
4
]. The ICP is an

iterative algorithm and consists of two steps. The first step is to find a correspondence between

target and source points, based on Euclidean distance between points. In the second step, the

updated version of result transformation is calculated using equation:

(1)

where
T
,
S
are target and source set of points,
N
s
is number of source points (equals number of

target points), and Rot, Trans are rotation and translation components of final transformation.

The updated version of final transformation in current iteration is based on close-form solu-

tion of mean square error problem. The classical approach used only one rigid or affine trans-

formation for whole data sets. In literature the description of disadvantages of the classical

• problem of finding global minimum of cost function depends on an initial guess of final

transformation,

• the algorithm is sensitive to improper correspondences,

• long time of computation—one of the most time-consuming operation is retrieving corres-

pondences.

Due to these disadvantages researchers proposed a lot of classical approach rectifications:•

• registration only subsets of points,

• improvement of finding correspondence problem,

• quantity measure of proper correspondence,

• elimination of improper correspondence,

• modification of computing minimum of cost function.

The standard ICP approach cannot be used to track surface of objects that change their

(2)

Search WWH ::

Custom Search