Division-Based Methods For Large Point Sets Registration
Loading...
Date
2016-03
Authors
Chen, Junfen
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Point sets registration is a key step for measuring the similarity between two point
sets and widely used in various fields such as computer vision, computer graphics,
medical image analysis, to name a few. The current devices can capture data with
great details as large point set. However, conventional registration methods slow down
dramatically as the size of the point set increased. In this thesis, three well-known
and among-best-performance point sets registration methods incorporating division
schemes are considered to study transforming conventional methods to efficiently deal
with large point sets registration. These methods are Iterative Closest Point (ICP),
Coherent Point Drift (CPD), and Gaussian mixture models based on thin-plate splines
(GMM-TPS). Firstly, a subset-ICP method is proposed based on streaming division
for rigid registration of large point sets. Instead of applying registration on the full
point set as the original ICP does, it recovers the rotation and translation using only the
correspondence between the pair of subsets. It iterates through all subset pairs until
convergent. Secondly, streaming division is incorporated to CPD-B method for nonrigid
registration of large point sets to recover the nonlinear displacement between
each point pairs among the two sets. Unlike the subset-ICP, it extends each subset
marginally to its neighbouring subset for robust point correspondence searching. The
registration results of all subset pairs are directly merged. A heuristic search is also proposed to tune the width parameter of the Gaussian kernel in the original CPD method,
namely S-CPD-B. Finally, bi-GMM-TPS is proposed as a two stage-based method for
large non-rigid point set registration. It employs a clustering scheme to obtain clusters
and registers the clusters’ centres using GMM-TPS to coarsely align the two full
point sets. It then finely registers all cluster pairs using GMM-TPS again. Extensive
experiments were conducted to validate the efficiency of the proposed methods on
the publicly available datasets including very large point set from USF database. The
experimental results demonstrated that the proposed methods are able to reduce the
computational cost as well as maintaining registration errors at comparable level with
the original methods. Three specific division-based registration methods are conceptually
summarised as a division-based multi-stage registration framework for handling
large non-rigid point sets registration.
Description
Keywords
Point sets registration