Object recognition using multi-view imaging
Author(s)
Wang, Yizhou
Type
Thesis
Abstract
Single view imaging data has been used in most previous research in computer vision and
image understanding and lots of techniques have been developed. Recently with the fast
development and dropping cost of multiple cameras, it has become possible to have many
more views to achieve image processing tasks. This thesis will consider how to use the
obtained multiple images in the application of target object recognition.
In this context, we present two algorithms for object recognition based on scale-
invariant feature points. The first is single view object recognition method (SOR), which
operates on single images and uses a chirality constraint to reduce the recognition errors
that arise when only a small number of feature points are matched. The procedure is
extended in the second multi-view object recognition algorithm (MOR) which operates on
a multi-view image sequence and, by tracking feature points using a dynamic programming
method in the plenoptic domain subject to the epipolar constraint, is able to fuse feature
point matches from all the available images, resulting in more robust recognition.
We evaluated these algorithms using a number of data sets of real images capturing
both indoor and outdoor scenes. We demonstrate that MOR is better than SOR particularly for noisy and low resolution images, and it is also able to recognize objects that are
partially occluded by combining it with some segmentation techniques.
image understanding and lots of techniques have been developed. Recently with the fast
development and dropping cost of multiple cameras, it has become possible to have many
more views to achieve image processing tasks. This thesis will consider how to use the
obtained multiple images in the application of target object recognition.
In this context, we present two algorithms for object recognition based on scale-
invariant feature points. The first is single view object recognition method (SOR), which
operates on single images and uses a chirality constraint to reduce the recognition errors
that arise when only a small number of feature points are matched. The procedure is
extended in the second multi-view object recognition algorithm (MOR) which operates on
a multi-view image sequence and, by tracking feature points using a dynamic programming
method in the plenoptic domain subject to the epipolar constraint, is able to fuse feature
point matches from all the available images, resulting in more robust recognition.
We evaluated these algorithms using a number of data sets of real images capturing
both indoor and outdoor scenes. We demonstrate that MOR is better than SOR particularly for noisy and low resolution images, and it is also able to recognize objects that are
partially occluded by combining it with some segmentation techniques.
Date Issued
2009
Date Awarded
2010-01
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Brookes, Mike
Creator
Wang, Yizhou
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)