Electronic Thesis and Dissertation Repository

Degree

Doctor of Philosophy

Program

Computer Science

Supervisor(s)

Dr. John Barron, Dr Steven Beauchemin

Abstract

We introduce a set of algorithms for registering, filtering and measuring the similarity of unorganized 3d point clouds, usually obtained from multiple views.

We contribute a method for computing the similarity between point clouds that represent closed surfaces, specifically segmented tumors from CT scans. We obtain watertight surfaces and utilize volumetric overlap to determine similarity in a volumetric way. This similarity measure is used to quantify treatment variability based on target volume segmentation both prior to and following radiotherapy planning stages.

We also contribute an algorithm for the drift-free registration of thin, non- rigid scans, where drift is the build-up of error caused by sequential pairwise registration, which is the alignment of each scan to its neighbor. We construct an average scan using mutual nearest neighbors, each scan is registered to this average scan, after which we update the average scan and continue this process until convergence. The use case herein is for merging scans of plants from multiple views and registering vascular scans together.

Our final contribution is a method for filtering noisy point clouds, specif- ically those constructed from merged depth maps as obtained from a range scanner or multiple view stereo (MVS), applying techniques that have been utilized in finding outliers in clustered data, but not in MVS. We utilize ker- nel density estimation to obtain a probability density function over the space of observed points, utilizing variable bandwidths based on the nature of the neighboring points, Mahalanobis and reachability distances that is more dis- criminative than a classical Mahalanobis distance-based metric.


Share

COinS