Using structure for feature localisation in images
Date
2012
Authors
Shaw, David Lawrence
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Techniques that are fast for matching in computer vision, such as the bag-of-features model, generally gain their speed by ignoring the relative structure of local elements. In comparison techniques that incorporate structure such as the constellation model typically involve an exhaustive analysis of all combinations that is computationally expensive. Efficient applications in matching the use of structure within the image, particularly perspective structure, has been inadequately treated.
This thesis explores the incorporation of geometric structure into image analysis techniques for efficient scene matching using methods derived from perspective geometry: planar properties, rectangles, square features and vanishing points. It develops and provides a series of techniques with the objective of real-time applications. The initial exploration into using geometry via polygonal properties of the environment as a basis for feature detectors is explored in the form of the Square Feature Detector (SFD). This detector looks for regions that exhibit square-like properties, using this as a low level feature and has the potential to complement the detectors currently available. The SFD shows advantages over leading point features (like the Scale-Invariant Feature Transform) for specific domain situations such as local baseline tracking during camera pans or robot rotations.
Perspective transformation is a weakness with the SFD, and led to the development of the Perspective Rectangle Detector (PRD). The PRD uses this perspective structure as the basis for a feature detector. This detector finds rectangles within the environments using the perspective structure of indoor scenes. These rectangles form a higher order feature aimed for use in both matching as well as localisation and navigation. Since the rectangles themselves are part of the environment they can be considered to represent structure in the intuitive sense, being constituent elements of the scene (doors, wall panels, and other rectangular objects). The Perspective Rectangle Detector contributes by combining the detection of rectangles under perspective with an emphasis on efficiency.
A Vanishing Point Detector was developed in conjunction with the rectangle detector, designed for use for similar real-time applications. This vanishing point detector is designed to take an "any-time" approach using estimated samples, allowing the detector to make increasingly accurate measurements over time while providing an approximate answer for real-time results. The benefit of this vanishing point detector over others is its interruptibility combined with the theoretically perfect accuracy given infinite time. Additionally, to help to further structure the features found, a simple but effective technique is presented for using the projective geometric property of the persistence of the order of angles between points under perspective projection. This Perspective Invariant Angle Order (PIANO) test is added to the bag-of-features matching technique to gain improvements in situations with multiple similar objects. The angle ordering is also applied and compared against existing visual codebook techniques for its potential as a viable alternative or for combination with existing methods. -- provided by Candidate.
Description
Keywords
Citation
Collections
Source
Type
Thesis (PhD)
Book Title
Entity type
Access Statement
Open Access
License Rights
Restricted until
Downloads
File
Description